supplementary information - Nature

SUPPLEMENTARY INFORMATION

doi:10.1038/nature10983

Supplementary Information for Curtis et al.: The genomic and transcriptomic architecture of 2,000 breast tumours reveals novel subgroups

Table of Contents Supplementary Methods

3

1

Biospecimen collection and ethical consent

3

2

Histopathological review

3

3

Experimental assays and orthogonal validation 3.1 Microarray data description . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Nucleic acid isolation and quality assessment . . . . . . . . . . . . . . . . . . . 3.3 DNA labelling and microarray hybridisations . . . . . . . . . . . . . . . . . . . 3.4 RNA labelling and microarray hybridisations . . . . . . . . . . . . . . . . . . . 3.5 Multiple Ligation-Dependent Probe Amplification (MPLA) assay . . . . . . . . . 3.6 Long-range PCR validation of germline CNVs and somatic homozygous deletions 3.7 TP53 sequencing and analysis . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

4 4 4 4 4 4 5 5

Gene expression analysis 4.1 Data pre-processing and normalisation 4.2 PAM50 classification . . . . . . . . . . 4.3 GENIUS classification . . . . . . . . . 4.4 Expression-based classification . . . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

5 5 6 7 7

4

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

5

Genotype calling

7

6

Inference of sample ethnicity

7

7

Confirmation of sample identity

8

8

Segmentation and copy number alteration calling using CBS 8.1 Normalisation of intensities . . . . . . . . . . . . . . . . . . . . . . . . . 8.2 Calling of copy number alterations and adjustment for tumour cellularity 8.3 Identification of germline CNVs . . . . . . . . . . . . . . . . . . . . . . 8.4 Gene-centric alterations . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

8 8 9 10 10

1

WWW.NATURE.COM/NATURE | 1

RESEARCH SUPPLEMENTARY INFORMATION

doi:10.1038/nature10983

8.5 8.6 8.7 9

Probe-level alterations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Measures of genomic instability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Identification of recurrent alterations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Segmentation and copy number alteration calling using HMM 9.1 Normalisation of intensities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.2 6-state HMM for segmentation and discrete copy number prediction in normals . 9.3 HMM-Dosage for segmentation and discrete copy number prediction in tumours . 9.4 Gene-centric alterations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . .

. . . .

. . . .

. . . .

. . . .

11 12 12

12 . 12 . 12 . 13 . 14

10 Comparison of segmentation methods based on MPLA

14

11 The copy number landscape of breast cancer 11.1 The germline CNV landscape . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 The somatic CNA landscape . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

16 16 17

12 eQTL analysis 12.1 Preprocessing of predictors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.2 Preprocessing of responses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.3 eQTL analysis of tumour gene expression profiles . . . . . . . . . . . . . . . . . . . . . . . 12.4 Effect of ranking expression values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.5 Definition of cis/trans windows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.6 Assessment of genotype calls in non-diploid regions and the impact of population stratification 12.7 eQTL analysis of normal gene expression profiles . . . . . . . . . . . . . . . . . . . . . . . 12.8 Trans-acting aberration hotspots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

17 18 18 19 20 21 21 22 23

13 Analysis of somatic copy number aberrations and gene expression 13.1 Subtype-specific analysis of copy number . . . . . . . . . . . . . . . . . . . . . 13.2 Cis-acting copy number and differential expression associations between groups . 13.3 Cis copy number and expression correlation analysis . . . . . . . . . . . . . . . 13.4 Genome-wide copy number and expression correlation analysis . . . . . . . . .

23 23 24 24 24

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

14 Cluster analysis 25 14.1 Integrative clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 14.2 K-Means clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 15 Detection of outlier expression profiles

29

16 Pathway analysis

29

17 Survival analysis

30

References

30

18 Supplementary Figures

34

19 Supplementary Table Legends

141

2



doi:10.1038/nature10983

Supplementary Methods 1

Biospecimen collection and ethical consent

We assembled a collection of over 2,000 clinically annotated primary fresh frozen breast cancer specimens and a subset of normals that passed initial selection criteria from tumour banks in the UK and Canada (for which a subset of 2,136 expression arrays and 2,477 Affymetrix SNP 6.0 arrays are reported on here). Both primary breast tumours, with linked pseudo-anonymised clinical data, and normal breast tissue were obtained with appropriate ethical approval from the relevant institutional review board. The METABRIC study protocol, detailing the molecular profiling methodology, was also approved by the ethics committees in Cambridge and Vancouver, the two sites responsible for the molecular analysis of the samples. As described below, tumours were primary invasive breast carcinomas for which clinical information could be categorically linked to DNA and RNA specimens. Initial quality control involved assessment of array quality and the flagging of mismatches between DNA and RNA using a novel eQTL-based approach (outlined below). Following the exclusion of cases on histopathological grounds (benign cases, those with ductal or lobular carcinoma in situ (DCIS, LCIS), or low tumour cellularity), incomplete clinical/pathological data (absence of ER status, grade, or tumour size), or apparently related individuals (based on genotype calls), paired DNA and RNA profiles were available from tumours derived from 997 female patients. DNA from adjacent normal breast tissue (or peripheral blood for some cases) was available from 485 samples, a subset of which match tumours in the discovery set. High quality RNA derived from adjacent normal tissue was available from 144 samples. The demographics, clinical, and pathological characteristics of the patients described in the discovery set are presented in Tables S1 and S2. A second cohort of 995 cases was later assembled for which matched DNA or RNA profiles or clinical information was not available at the time of the initial analyses, and these included low cellularity tumours, DCIS, and three benign cases. This cohort represents a validation set, which was employed for the purposes of testing the reproducibility of the integrative clusters and clinical outcome associations. Genotype analysis for the full set of cases subsequently revealed that eight individuals were represented both in the discovery and validation set (MB: 0667/0025, 0546/0326, 0327/0547, 0549/0329, 0559/0335, 0573/0355, 0408/0407, 0432/0433), and four were represented twice in the validation set (MB: 0110/0196, 0552/0330, 6213/6206, 2820/2720), but were supplied as unique accessions by the tumour bank. These sample pairs represent multiple primary tumours from the same individual and different sections of the same tumour. These comprise only 1.2% of samples in the validation set and do not alter the conclusions from these analyses. The demographics, clinical, and pathological characteristics of the patients described in the validation cohort are presented in Tables S1 and S3.

2

Histopathological review

The frozen tissue sections from which nucleic acids were isolated were subject to expert histopathological review to assess the presence of invasive tumour, pre-malignant or benign changes, tumour cellularity, and lymphocytic infiltration in specific subgroups (Table S45). Tumour cellularity was scored visually in a semiquantitative fashion on three sections (taken at the beginning, middle, and end of cryosectioning) per tumour for UK samples and two sections (taken at the beginning and end of cryosectioning) per tumour for Canadian samples, where cellularity values were binned so that ‘low cellularity’ corresponds to samples with 70% tumour DNA were considered to have ‘high cellularity’. As noted above, samples that were classified as ductal or lobular carcinoma in situ (DCIS, LCIS) or as benign were excluded from the discovery set.

3



doi:10.1038/nature10983

3 3.1

Experimental assays and orthogonal validation Microarray data description

Matched DNA and RNA were extracted from each specimen and subject to copy number and genotype analysis on the Affymetrix SNP 6.0 platform and transcriptional profiling on the Illumina HT-12 v3 platform (Illumina_Human_WG-v3). A flow chart depicting the main analytical approaches that utilise these data is presented in Figure S1 (supporting files are available at: http://www.compbio.group.cam.ac.uk/resources.html). The associated genotype and expression data have been deposited at the European Genome-Phenome Archive (EGA, http://www.ebi.ac.uk/ega/), which is hosted by the European Bioinformatics Institute, under accession number EGAS00000000083.

3.2

Nucleic acid isolation and quality assessment

For UK samples, DNA and RNA were extracted from ten 30 µm sections each from fresh frozen tumours using the DNeasy Blood and Tissue Kit and the miRNeasy Kit (Qiagen, Crawley, UK) on the QIAcube (Qiagen) according to the manufacturer’s instructions. For Canadian samples, DNA and RNA were extracted from 10 eight µm sections each from fresh frozen tumours using the MagAttract DNA Mini M48 Kit and miRNeasy 96 Kit (Qiagen) manually in a 96-well format according to the manufacturer’s instructions. Nucleic acids were quantified with a NanoDrop ND-8000 spectrophotometer (NanoDrop Technologies, Wilmington, DE, USA) and quality assessed by agarose gel electrophoresis. RNA quality was also assessed using the Agilent 2100 Bioanalyser Nanochip (Agilent Technologies, Wokingham, UK). Tumour samples for which the RNA had an RNA Integrity Number (RIN) > 7 were hybridized to expression arrays, whereas a less stringent RIN > 5 was required for normal RNA. Random DNA and RNA samples were selected for genotyping purposes to ensure sample uniqueness using the AmpF l STRIdentifiler PCR Amplification Kit (Applied Biosystems, Foster City, CA, USA).

3.3

DNA labelling and microarray hybridisations

DNA was hybridized to Affymetrix SNP 6.0 arrays per the manufacturer’s instructions (Affymetrix, Santa Clara, CA) at AROS Applied Biotechnology (Aarhus, Denmark). Samples that met the quality control criteria established by AROS and suggested by the Affymetrix Genotyping Console v2.1 were subject to further in-house quality assessment (as described below).

3.4

RNA labelling and microarray hybridisations

Total RNA was used to generate biotin-labelled cRNA using the Illumina Totalprep RNA amplification kit (Ambion, Warrington, UK) and hybridized onto Illumina Human HT-12 v3 Expression Beadchips per the manufacturer’s instructions and was scanned on the Illumina BeadArray Reader.

3.5

Multiple Ligation-Dependent Probe Amplification (MPLA) assay

Multiplex ligation-dependent probe amplification (MLPA) was used to validate copy number changes observed in the SNP-CGH data. Two commercially available kits were used, namely, the SALSA MLPA P005 Human Chromosomal Aberration-1 panel and the SALSA MLPA P078 Breast tumour panel. Assays were performed according to the manufacturer’s instructions (MRC-Holland, Amsterdam, the Netherlands) on 40 tumours. 4



doi:10.1038/nature10983

3.6

Long-range PCR validation of germline CNVs and somatic homozygous deletions

Primers were designed for copy number neutral regions adjacent to the homozygous deletion using Primer 3 software (http://frodo.wi.mit.edu/primer3) (Table S10). PCR reactions were performed using the SequalPrep Long PCR kit and Enhancer B Buffer according to the manufacturer’s instructions (Invitrogen, Paisley, UK). Prior to sequencing, LR-PCR products were confirmed by gel electrophoresis. Several CNVs (and somatic homozygous deletions) were validated to nucleotide resolution in DNA from the tumours and/or MN tissue via PCR and sequencing (Table S9) as illustrated for a few exemplary cases in Figure S3.

3.7

TP53 sequencing and analysis

As TP53 mutations also contribute to GI, we surveyed the TP53 mutational spectrum of 820 cases from our cohort by Sanger sequencing (Table S2). Mutations in the 11 TP53 exons were examined as previously described [1] (with the exception that exon 7 was sequenced in only one direction). Mutations were scored using Mutation Surveyor software (v3.23, SoftGenetics, LLC, State College, PA, USA). To increase the mutation detection sensitivity, we applied the following modifications to the default settings: 1.) Data were analysed with unidirectional parameters instead of bidirectional parameters. This allows the detection of mutations in areas of high noise; 2.) The "Check 2D Small Peaks (Mosaic)" box was checked. This option enables the software to display potential somatic mutations which often have low intensity; 3.) The ‘dropping factor’ was decreased from 0.2 to 0.1. The dropping factor measures the level of decrease in the sample’s normal allele. By decreasing the dropping factor from 0.2 to 0.1, the normal allele concentration can drop less before a mutation would be detected. 4.) The Mutation Height was decreased from 500 to 300. All the mutations called by Mutation Surveyor were inspected visually to eliminate false positives and false negatives, but note that this automated method is likely to under call frameshift mutations and thus the mutation frequency reported is likely to be an underestimate. Splice site mutations were not included, but would result in a negligible increase in the number of affected cases (2 cases). Sequences that failed in one direction were scored as unidirectional. Mutations were then manually confirmed using Lasergene’s SeqMan Pro (DNASTAR, Inc. Madison, WI, USA). In total, 18,075 Mutation Surveyor reports covering 820/997 cases were parsed and projected onto genomic coordinates using RefSeq accession NC_000017 region 7512445-7531642. All variants were filtered against dbSNP (v129) and amino acid annotations were obtained using MutationAssessor (http://mutationassessor.org). All mutations required support from both directions to be called. A small subset of the mutations were validated by a second PCR reaction and Sanger sequencing. The overall frequency of TP53 mutation was ∼12% (n = 99; 69 missense, 30 truncating) with 84% of mutations present in histologically high-grade (Grade 3) tumours predominantly of the Basal-like subgroup (34% of cases, n = 33) (Figure S5). As such, the genome-wide copy number profiles of TP53 mutated cases reflect the distinct pattern of the Basal-like tumours (Figures S32, S33), as has been noted by others [2].

4 4.1

Gene expression analysis Data pre-processing and normalisation

A custom R [3] script was written to process each BeadChip once scanning was complete and raw data were available. The script relied on existing functionality within the beadarray package [4], including the generation of quality assessment (QA) information and adjustment for spatial artefacts with the BASH tool [5]. Once all 12 arrays on the chip had been processed, the bead-level data were summarized, yielding

5



doi:10.1038/nature10983

a series of 48,803 × 12 matrices of log2 intensities, standard errors, and number of observations. After all chips had been processed, the summarized matrices were combined. Potential outlier arrays were removed by considering the bead-level QA scores derived using the control probes on each array. We resisted the temptation to use methods that compare array intensities and distributions between arrays to allow for the possibility that these represent biologically interesting differences. Arrays with a P95 scan metric less than 200 were excluded as they were considered to have failed hybridisation. These arrays were then removed from the dataset so as to not influence the following QA. A multivariate outlier testing procedure [6] implemented in the arrayMvout Bioconductor package was used to identify arrays with poor quality based on the bead-level QA scores of all arrays. Additionally, this approach was applied to bead-level QA information for each location separately. All arrays that remained after this three-step procedure were retained in the subsequent analysis. We relied on the comprehensive re-annotation of the Illumina HT-12 v3 platform [7] as we have noticed the sub-optimal performance of certain probes, and therefore wanted to avoid the influence of such probes on the normalisation. To do so, we first generated a list of suitable probes based on the following criteria: must be a perfect genomic match, does not target the sex chromosomes, must not contain a SNP, must be 30 matching and without multiple genomic matches, must not contain a polyG tail at the end of probe sequence, GC content must be between 38% and 64%, and the probe must not target genes from the PAM50 list. As we did not note any differences in the target distributions from different tumour banks, a single target distribution was generated using the list of suitable probes. ER-positive and ER-negative samples were quantile normalised separately and averaged to obtain the final target distribution. Each array was then normalised to the target by quantile normalising probes belonging to the target distribution, while the values for the remaining probes were obtained by interpolation using the weighted normalised intensities of the target distribution probes with most similar intensities prior to normalisation. A linear model was then fit using the limma [8] Bioconductor package to remove any batch effects associated with the position of an array on the Illumina BeadChip. Differential expression analysis for contrasts of interest was performed on a subset of the log2 normalised intensity matrix that excluded any probes mapping to non-transcribed regions. This approach has recently been shown to improve the sensitivity of a differential expression analysis [7]. A probe-wise linear model was fit to the data using limma with coefficients estimated for particular clinical variables of interest (for example, ER-positive or ER-negative cases). Differences in expression level arising from characteristics of the different tumour banks were accounted for by incorporating a coefficient for tumour-bank in the linear model. Furthermore, a weight matrix, defined to be the square root of the number of observations on each array, was used to down-weight arrays with fewer observations in the model. After empirical Bayes’ moderation of the variance, adjusted P-values, log-ratios, and log-odds were used to assess the evidence for differential expression in each contrast of interest.

4.2

PAM50 classification

Samples were classified into the five intrinsic subtypes [9, 10] based on PAM50 [11]. As we have previously noted probe annotation to be an important consideration in sample classification using microarray data [12], the PAM50 gene-list was refined so that only genes for which a corresponding probe with perfect annotation on the Illumina HT-12 v3 BeadChip [7] were used for classification. As a result, BAG1 and TMEM45B were not included in the classification. For genes with more than one probe, probes were selected on the basis of their annotation. For example, probes containing a SNP were avoided as were those which target introns, have secondary targets or mismatches, lie in repeat-masked regions, or target the wrong genomic strand. As previously recommended [13], all probes were median centred prior to classification. Due to the

6



doi:10.1038/nature10983

imbalance in ER status, we defined 100 random reference distributions consisting of all ER-negative samples and randomly selected ER-positive samples during the median centring step. This resulted in 100 different classifications and the final subtype calls were derived by taking a consensus across all 100 trials. Samples were then assigned to one of the five intrinsic subtypes using the Spearman correlation to the published centroids and the transformed intensities, where samples with correlations < 0.1 for all subtypes were not classified (NC) (0 samples in the discovery set, 6 samples in the validation set).

4.3

GENIUS classification

Samples were classified into the four Gene Expression progNostic Index Using Subtypes (GENIUS) subgroups (HER2+, ER-/HER2-, ER+/HER2- high proliferation, ER+/HER2- low proliferation) [14] using the genefu [15] package in Bioconductor.

4.4

Expression-based classification

ER, PR, and HER2 status were classified based on their empirical expression distributions using a 2component Gaussian mixture model as implemented in MCLUST [16, 17] in an approach similar to that described by Lehmann et al. [18].

5

Genotype calling

Affymetrix SNP 6.0 arrays were pre-processed and genotyped using the SNP-RMA algorithm [19], available in the crlmm [20] Bioconductor package. The goal of pre-processing is to convert raw intensity values into quantities proportional to the amount of DNA in the target sample associated with each of the alleles, A and B, for each SNP. Briefly, feature intensities were corrected for fragment length and sequence effects, followed by quantile normalisation to a predefined reference distribution. Intensities were then summarized by median polish, resulting in a single value for each allele. A mixture model was then used to adjust for remaining fragment length and intensity-dependent biases on the log-ratio of the summarized intensities. Samples with a signal-to-noise ratio < 5 were flagged in downstream analyses.

6

Inference of sample ethnicity

Genotype calls for each sample were projected onto the HapMap principal component loadings for Affymetrix in order to infer sample ethnicity. The HapMap PCA loadings for Affymetrix SNP probes (snpload.aff.map) were downloaded from: http://www.stats.ox.ac.uk/ davison/software/shellfish/shellfish.php and read into R. Probe annotations and allele frequencies for each of the four HapMap populations (CEPH (CEU); European, Yoruban (YRI); African, Han Chinese (CHB), and Japanese (JPT)) were extracted from the Affymetrix GenomeWideSNP_6.na28.annot.csv file. In total, 168,984 SNPs in the PCA loadings file map to the SNP 6.0 platform and were used in this analysis. For all SNPs in the loading file, the positive strand was taken as a reference. Assuming a coding of 0, 1, and 2 for the SNPs and that they occur in Hardy-Weinberg equilibrium at a ratio of p2 :2p(1-p):(1-p)2 , the relevant allele proportions were extracted and a population score computed for each SNP. Multiplying these scores by the PCA loadings yielded the principal component scores and for each individual, the relevant entries were again multiplied by the PCA loadings. Plotting these data revealed that the two Asian populations were largely overlapping. As such, three main clusters were defined (Africa, Asia, and Europe), and the distance from the cluster to each individual was calculated. Ethnicity 7



doi:10.1038/nature10983

calls were made based on the minimal distance between a given sample and population cluster (Figure S7, Table S16).

7

Confirmation of sample identity

We employed a novel eQTL-like approach to ensure that no plating errors had occurred, and that when we claim that the same sample was run on the expression and genotyping platforms we were indeed mapping between the two correctly. This approach we have termed BeadArray Diagnostic for Genotype and Expression Relationships (BADGER) (Lynch et al., manuscript in preparation). Briefly, a panel of 383 strong (-log10 P-value > 15) relationships between measurements of log-expression and reported genotypes was collated from the data. Many of these will be technical associations (e.g. the expression probe covers a SNP), and not genuine eQTLs, but this is irrelevant for this purpose. For each expression array, using this panel and these relationships, a predicted genotype profile was generated. The predicted profile was then compared to each observed profile for those 383 SNPs to ensure that the mappings were correct. BADGER revealed a small number of plating errors that have now been corrected. For all but two of the samples, the match between predicted and observed genotype was exceptional, with the remaining discrepancies probably due to mixture of relatively noisy expression data and possibly the non-diploid nature of the tumour DNA.

8

Segmentation and copy number alteration calling using CBS

Tumour sample copy number log2 ratios were segmented with two analytical methods, namely circular binary segmentation (CBS) [21] and an adapted hidden Markov model (HMM) [22] to provide contrasts between sensitivity and specificity (see Comparison of segmentation methods based on MPLA), and we present the results based on both methods. The maintext figures are based on the CBS-derived copy number profiles with the exception of the outlier expression analysis (Figure 2). Where applicable, the HMM-based results are presented in the Supplementary Information. As outlined below the copy number data were rigorously validated on a subset of cases using several independent techniques (see Methods, Figure S3, Table S9, S10), and allowed for estimation of the Type I and Type II error rates for both segmentation methods (Figure S4, Table S11).

8.1

Normalisation of intensities

Affymetrix SNP 6.0 arrays were pre-processed for copy number segmentation using aroma.affymetrix. Both tumour and normal samples were independently normalised using the single array method referred to as Copy-number estimation using Robust Multichip Analysis (CRMAv2) [23], as was a publicly available SNP 6.0 dataset consisting of 270 HapMap individuals. We applied the default settings using the following tags: ACC,ra,-XY,BPN,-XY,AVG,A+B,FLN,-XY. For each sample, allelic-crosstalk calibration, probe sequence effects normalisation, probe-level summarisation, and PCR fragment length normalisation were performed in order to obtain log2 intensity values for total copy number estimation. The following annotation files were used (Feb 14, 2008): • Chip definition file (GenomeWideSNP_6,Full.cdf) • Unit fragment-length (GenomeWideSNP_6,Full,na24,HB20080214.ufl) • Unit genome position (GenomeWideSNP_6,Full,na24,HB20080214.ugp) 8



doi:10.1038/nature10983

Following these steps, probes were sorted according to their genomic position, replicate probes were summarized by their median value, and missing values (produced by negative intensities in the normalisation) were imputed using the loess procedure included in the snapCGH [24] Bioconductor package. Various QC metrics were employed to assess the SNP 6.0 copy number data, including the normalised unscaled standard error, relative log expression, and signal-to-noise ratio of the log2 intensity data. Samples for which any of these metrics were flagged were visually inspected in order to determine whether to retain them in the analysis. Two pooled references were created, one using the median intensities across the HapMap individuals and another for the normals and tumours, using the median intensity values from a set of 473 normals. Different pools based on gender were created for the X chromosome. Next, log2 ratios were generated for the HapMap samples by subtracting the pooled value from the log2 intensities. Similarly, log2 ratios were obtained for the 473 normals using the corresponding pool. For the 997 tumour samples, two data sets were produced: one using the normal pool as the reference for all the tumours and another using the matched normal for each tumour when available, and the normal pool for the remainder. A similar approach was taken for the validation set. Following genotype analysis, we noted that a small number (13) of the normals were not correctly matched to tumours in the validation set, but this would have a nominal impact on the copy number calls due to the strict criteria for calling common CNVs (outlined below). The HapMap and normal datasets were employed to estimate the frequency of germline CNVs in the cohort, while the tumour samples were used for estimating somatic CNAs. After computing the log2 ratios for each probe, samples were segmented using the circular binary segmentation (CBS) algorithm [21] implemented in the DNAcopy [25] Bioconductor package and alterations were called as described below.

8.2

Calling of copy number alterations and adjustment for tumour cellularity

DNAcopy with prior smoothing and default parameters was run on the set of 473 normals and 270 HapMap samples. Although Ostrovnaya et al. [26] noted that not smoothing facilitated the detection of CNVs, we found that this led to higher error rates (false positives). For calling alterations, thresholds were set based on the median of the log2 ratio for each array +2σ or -2.5σ (where σ is the standard deviation of the log2 ratio for each array) for calling gains and losses, respectively. These asymmetric thresholds were based on the assumption that the expected log2 ratio for a single copy gain is log2 (3/2), which is smaller in absolute value than the expected log2 ratio for a single copy loss (log2 (1/2)). This was also confirmed by an independent analysis of the MLPA data (see below). In order to discover common CNVs, at least five samples in a particular population were required to exhibit an alteration in the same probe. We did not put any restrictions on the length or number of probes required to call a CNV, although implicitly there is a restriction on the minimum length of a segment required by DNAcopy (2 probes). Consecutive probes were merged to call copy number variable regions and separate lists were obtained for the different HapMap populations (CEPH, YOR, CH-JA) and the 473 normals. Finally, the three HapMap populations were merged with the set of 473 normals to generate a list of CNVs present in these populations (HapMap+Normals CNV list). For tumour samples, the data were smoothed and DNAcopy was run with default parameters. We then applied the MergeLevels algorithm [27] to the segmented data. It can generally be assumed that there is a mixture of normal (benign) and tumour cells present within a given sample. In order to remove the dependence between cellularity and the proportion of alterations, we employed different thresholds for calling alterations and high-levels events according to the cellularity of each sample, resulting in the following somatic copy number states: KCNA = {HOMD, HET D, NEUT, GAIN, AMP} .

(1)

9



doi:10.1038/nature10983

For ‘high cellularity’ tumours, the median of the log2 ratio +2σ or +6σ was computed for the 50% of the central probes (ordered by their log ratios) to call gains and amplifications, respectively. For losses, the median of the log2 ratio -2.5σ or -7σ was used to call heterozygous and homozygous losses, respectively. For ‘moderate cellularity’ tumours, the median of the log2 ratio of each array +2σ or +6σ was computed for the 45% of the central probes and used to call gains and amplifications, respectively. For losses, the median of the log2 ratios -2.5σ or -7σ was used to call heterozygous and homozygous losses, respectively.

8.3

Identification of germline CNVs

For the tumour samples, any segmented mean that fell within a region included in the HapMap+Normals CNV list was flagged as an inherited CNV, yielding the following germline copy number states: KCNV = {CNV LOS S , CNVGAIN}

(2)

In order to remove all possible CNVs, the frequencies of CNAs in the tumour samples were obtained after removing these CNVs from the data normalised with the pool and the corresponding matched normals, when available. In order to retain all possible CNVs, the frequencies of CNVs in the tumour samples were obtained with these CNVs flagged in the data normalised only with the pool. We further compared the distributions of CNVs derived from tumour-adjacent normal breast tissue to those obtained from constitutional blood DNA in a separate cohort of 1,999 breast cancer patients (unpublished data), and found that 92.8% of CNVs identified in these normals were also called in blood, suggesting that adjacent normal breast tissue is a suitable surrogate for germline CNV profiling. The list of CNV frequencies, gene-centric Ensembl-based CNV mappings, and population distributions are presented in Tables S12-S15, and a plot illustrating that the CNV landscape was the same irrespective of the intrinsic subtype is presented in Figure S6.

8.4

Gene-centric alterations

In order to identify the genes and features that were altered by copy number aberrations, we searched for the overlap of segments with gene regions. The complete set of gene annotations and coordinates given by Ensembl 54 (hg18) that correspond to Illumina HT-12 v3 probes was employed, resulting in 30,566 features. Two patient-by-gene copy number matrices were generated to capture different perspectives of gene alterations. The call matrix, C, was populated with values representing discrete copy number states, whereas the log-ratio matrix, L, contains the segmented mean. For each patient p ∈ P and each gene g ∈ G, we identified segments s that overlap g and assign C(p, g) with the copy number state of s and L(p, g) with the mean log-ratio of the probes in s. If gene g overlaps or is broken by a set of segments, S = s1 , . . ., sk , where k ≥2, we assign L(p, g) with the maximal log-ratio (Equation 3). For C(p, g), the copy number state of the segment with maximal severity was assigned based on the relationship defined in (Equations 4 and 5), where ties were broken in samples exhibiting both a loss and gain according to the maximal absolute value of the segmented mean.

10



doi:10.1038/nature10983

L(p, g) =

X

X length(s) ∗ meanLogR(s) / length(s)

s∈S

(3)

s∈S

C(p, g) = CN state argmax{severity(CN state(s)) s∈S severity = ‘NEUT ’ < ‘HET D’ < ‘HOMD’, ‘NEUT ’ < ‘GAIN’ < ‘AMP’, ‘CNV NEUT ’ < ‘CNV LOS S ’ = ‘CNVGAIN’

(4)

(5)

Note that somatic and germline events were treated separately, but that in some cases both types of events can occur within different segments of the same gene. On occasion, an expression probe may fall in the middle of a copy number breakpoint for a particular sample. In this case there is no way to interpolate the value, so the copy number value for the probe is deemed missing. The discrete call matrix was subsequently used in ANOVA to identify copy number events significantly associated with changes in gene expression, the top associations of which were selected as features for integrative clustering as described below. As the call matrix only allows for one copy number state per gene, when multiple states are in fact possible, we also report the frequency of samples showing all possible levels of alteration within a gene for the purpose of understanding the entire spectrum of gene copy number alterations in this cohort. Thus a sample can exhibit both a gain and neutral segment spanning the same gene or, alternatively, a region of loss and region of gain within a gene. Here, a sample may contain a region of loss and a region of gain in a particular gene, but for a sample to be considered neutral it must be so for the entire length of the gene. These frequencies were computed for all tumours, as well as by ER status, PAM50 subtype, and the novel integrative cluster-based subgroups defined in this study. For these summaries, we extended the feature set to include ncRNAs and other genomic regions not necessarily targeted by the Illumina HT-12 v3 array by using the complete set of annotations and coordinates given by Ensembl 54 (hg18), with the exclusion of ChrNT and ChrMT genes, resulting in 36,656 features (36,209 when ChrY is excluded). The frequency with which a gene (feature) exhibited a copy number aberration in a particular subtype is reported in Table S4 and the frequency with which a gene was broken by a copy number event is reported in Table S6. Subtype-specific copy number profiles for the PAM50 intrinsic subtypes are plotted in Figure S32. As described below, we also computed a number of summaries of the copy number data that were not restricted to a gene scaffold.

8.5

Probe-level alterations

We computed the frequency of alterations for each probe in all tumours and according to ER status, PAM50 subtype and the novel integrative subgroups defined in this study. Frequencies of alteration were computed for germline CNV states: KCNV = {CNV LOS S , CNVGAIN} (6) and for each of the 5 discrete somatic states: KCNA = {HOMD, HET D, NEUT, GAIN, AMP}

(7)

in addition to the collapsed summaries of alterations represented by loss (HOMD & HETD), neutral (NEUT), and gain (GAIN & AMP). Several probe-level plots of putative homozygous deletions are illustrated in Figure S17.

11



doi:10.1038/nature10983

8.6

Measures of genomic instability

The extent of genomic instability (GI) was computed for each sample using four related measures summarized from the CBS segmented data, namely the area under the segmented mean (area), proportion of probes altered (proportion), the proportion of genome altered (proportion genome), and the sum of changes in altered segmented means divided by the sum of all changes in segment means (e.g. the ratio of the mean change in log-ratio due to alterations relative to variability) (jump). These indices were computed for each sample and each chromosome band and clustering was performed jointly on the proportion of genome altered and jump measure using the Euclidean distance metric and Ward’s method in order to identify subgroups partitioned by genomic instability (Figure S26).

8.7

Identification of recurrent alterations

We employed the method of Rouveirol et al. [28] to compute the minimal common region of alteration for CNAs and CNVs present in the tumours according to ER status and PAM50 subtypes (Tables S5, S12). Due to the difference in their inherent structure, for the CNAs a minimum frequency of 0.05 and 3 probes was required, while for the CNVs no minimum length was imposed and the same minimum frequency was required. Regions of gains and losses were computed independently and were annotated for gene content based on Ensembl 54 (hg18).

9 9.1

Segmentation and copy number alteration calling using HMM Normalisation of intensities

The Affymetrix SNP 6.0 arrays for both tumour and normal samples were independently normalised using the single array method CRMAv2, implemented in aroma.affymetrix [23], as described above for CBS. Log-ratios were computed for both tumours and normals by normalising each array independently against a reference. Due to the incomplete pairing of tumours and matched normals, we generated a pooled reference from all the available normals. However, because we are also interested in the identification of CNVs in addition to CNAs in the tumours, using a simple pooled reference could potentially result in the ‘subtraction’ of frequent (> 50%) copy number events in the normals from the tumours. Therefore, we generated a ‘masked’ reference from the normal dataset that was free of inherent copy number polymorphisms, while accounting for systematic biases. Starting with the CRMAv2 normalised intensities for the normal samples, we computed log-ratios using a randomly permuted reference. Probe-level discrete copy number calls were then made made using a 6-state hidden Markov model (described below) such that for each sample, probes that were predicted to span regions of gain or loss were identified. Next, we retrieved the normalised intensities and masked out the probes identified in the previous step. Finally, we computed the median (pooling) intensity value for each probe across all samples using only the remaining values after masking, resulting in a vector of median values, which makes up the masked reference. For HMM, the masked reference improved the analysis of germline CNVs in the tumours by serving as a base-line vector of intensities for log-ratio calculations.

9.2

6-state HMM for segmentation and discrete copy number prediction in normals

We applied a 6-state HMM to the set of normal samples in order to detect germline CNVs by returning probe-level and segment copy number. The 6-state HMM is a modified version of CNA-HMMer [22] that

12



doi:10.1038/nature10983

has been adapted to analysing high-density genotyping arrays such as the Affymetrix SNP 6.0 platform. Another extension is the inclusion of additional copy number states, which offers a more intuitive interpretation of DNA dosage in cancer. The new discrete copy number state space included neutral (NEUT), homozygous (HOMD) and hemizygous (HETD) deletions, gain (GAIN), amplification, (AMP), and high-level amplification (HLAMP). KCNA = {HOMD, HET D, NEUT, GAIN, AMP, HLAMP}

9.3

(8)

HMM-Dosage for segmentation and discrete copy number prediction in tumours

We performed copy number analysis on the 997 tumour samples using HMM-Dosage (Ha et al., manuscript in preparation; http://compbio.bccrc.ca). The algorithm was designed to detect and distinguish the complete set of somatic and germline copy number events in cancer genomes interrogated by SNP array data. This model extends the 6-state HMM (above) by using 5 additional CNV states to represent the analogous copy number status KCNV = {CNV HOMD, CNV HET D, CNVGAIN, CNV AMP, CNV HLAMP}.

(9)

The model performs segmentation on log-ratios of intensity data and assigns a discrete copy number states kt ∈ {KCNA , KCNV } to each latent variable, Zt , for all probes t ∈ {1 . . . T }. Log-ratios, Y = {y1 , . . . , yT }, are observed values modeled using Student-t densities in the emission component of the HMM, conditional on Zt = kt . Spatial information is captured using a non-stationary transition matrix, At , that encodes positionspecific probabilities of CNVs and CNAs at each probe t. In order to distinguish between somatic copy number alterations (CNAs) and germline copy number variants (CNVs), HMM-Dosage probabilistically incorporates CNV information as a prior to the transition matrix. The CNV prior was computed as probe-level CNV frequencies by the weighted averaging of two datasets: the set of normal samples and an external dataset of 450 HapMap normal samples whose CNVs were predicted by Conrad et al. [29]. Because the HapMap dataset contains males, we excluded the X chromosome frequencies and instead used the X chromosome frequencies from the set of normals (all of which were female). The parameters of the Student’s t distributions are unobserved and estimated using the expectation maximisation (EM) algorithm for each sample, independently. Initial parameters for the KCNA states of the Student’s t mixture were empirically determined using 45 SNP 6.0 breast cancer cell line samples in the COSMIC [30] dataset. First, independently for each cell-line sample, we fit the log-ratios to a 6-state Gaussian mixture model using EM. Subsequently, the converged Gaussian mixture parameters were used as initial parameters for fitting a 6-state Student’s t mixtures in the 6-state HMM (again, using EM). Finally, we averaged the converged Student-t parameters for each state across the 45 cell lines. These averaged parameters became the initial KCNA parameters to Student’s t emission of HMM-Dosage. Initial parameters for KCNV were set by hand to theoretical values log2 [0.5, 1, 3, 5, 7] /2 . The optimal state sequence, Z1:T , which represents the copy number prediction for each probe, was computed using the Viterbi algorithm. HMM-Dosage was tuned to be conservative in calling CNVs so as to avoid misclassifying CNA calls. Therefore, the results were post-processed by comparison against the CNV map (derived from the set of normals and 270 HapMap individuals) such that CNA segments that had at least 25% reciprocal genomic overlap with a segment of the same copy number state in the CNV map were converted to the germline state.

13



doi:10.1038/nature10983

9.4

Gene-centric alterations

In order to identify the genes that were altered by copy number changes, we searched for the overlap of segments with gene regions. The gene annotations and coordinates are given by Ensembl 54 (hg18) and the analysis was restricted to those that are protein-coding and overlap with the annotations for the Illumina HT-12 v3 gene expression arrays, resulting in 18,733 genes. We generated two patient-by-gene copy number matrices to capture two perspectives of gene alterations. The first is the call matrix, C, which is populated with values representing discrete copy number state calls for each patient. The second is the log-ratio matrix, L, which contains segment median log-ratios. For each patient p ∈ P and each gene g ∈ G, we identify segments s that overlap with g and assign C(p, g) with the copy number state of s and L(p, g) with the median log-ratio of the probes in s. If gene g overlaps or is broken by a set of segments, S = s1 , . . ., sk , where k ≥2, we assign L(p, g) and C(p, g) with the weighted sum of the median LogRs (Equation 10) and the copy number state of the segment with maximal severity based on the relationship denoted in Equations 11 and 12. L(p, g) =

X

X length(s) ∗ medianLogR(s) / length(s)

s∈S

(10)

s∈S

C(p, g) = CN state argmax{severity(CN state(s)) s∈S severity = (‘NEUT ’, 0), (‘HOMD’, 8), (‘HET D’, 6), (‘GAIN’, 5), (‘AMP’, 6), (‘HLAMP’, 8), (‘CNV HOMD’, 4), (‘CNV HET D’, 3), (CNVGAIN’, 2), (‘CNV AMP’, 3), (‘CNV HLAMP’, 4)

(11)

(12)

The frequency with which a gene exhibited a copy number aberration in a particular subtype is reported in Tables S7 and the frequency with which a gene was broken by a copy number event is reported in Table S8. Subtype-specific copy number profiles for the PAM50 and integrative subtypes are plotted in Figures S33 and S28, respectively.

10

Comparison of segmentation methods based on MPLA

We validated the accuracy of both segmentation methods using the MLPA data, which assayed a panel of 90 commonly aberrant cancer loci in 40 tumours. MLPA produces a copy number estimate for each probe (gene) according to the following 6 states: homozygous loss [HOLoss] (0 copies), heterozygous loss [LOH] (1 copy), neutral [Normal] (2 copies), gain [Gain] (3 copies), amplification [Amplification] (more than 3 copies), and ambiguous (borderline value). Note that probes classified as ambiguous by MLPA were not included in the computations. MLPA probes were aligned with the genome using the exonerate program and the resulting coordinates were used to extract the corresponding CBS and HMM copy number states for each of the 90 × 40 MLPA values. The most extreme value was selected when multiple states were present in a given region in accordance with their respective call matrices, C. The following definitions were employed for these analyses: • Sensitivity: TP / (TP + FN)

14



doi:10.1038/nature10983

• Specificity: TN / (TN + FP) • Positive predictive value (PPV): TP / (TP + FP) • Negative predictive value (NPV): TN / (TN + FN) where TP = number of true positives, FN = number of false negatives, TN = number of true negatives, and FP = number of false positives. We used different definitions of TP, TN, FP, FN to assess the ability of the two copy number segmentation methods to detect different types of alterations. The notation ([Gain],[G,L]) indicates the number of probes with a gain in MLPA and a gain or loss according to the segmentation method: 1. Global: • TP = ([Amplification],[AMP]) + ([Gain], [GAIN]) + ([LOH, HETD]) + ([HOLoss], [HOMD]) • FN = ([Amplification], [GAIN, NEUT, HETD, HOMD]) + ([Gain], [AMP. NEUT, HETD, HOMD]) + ([LOH], [AMP, GAIN, NEUT, HOMD]) + ([HOLoss], [AMP, GAIN, NEUT, HETD]) • FP = ([Normal], [AMP, GAIN, HETD, HOMD]) • TN = ([Normal], [NEUT]) 2. High Level: • TP = ([Amplification],[AMP]) + ([HOLoss], [HOMD]) • FN = ([Amplification],[GAIN, NEUT, HETD]) + ([HOLoss],[GAIN, NEUT, HETD]) • FP = ([Gain, Normal, LOH, HOLoss],[AMP]) + ([Amplification, Gain, Normal, LOH],[HOMD]) • TN = ([Gain],[GAIN, NEUT, HETD]) + ([NEUT], [GAIN, NEUT, HETD]) + ([LOH], [GAIN, NEUT, HETD]) 3. Amplifications: • TP = ([Amplification],[AMP]) • FN = ([Amplification],[GAIN, NEUT, HETD, HOMD]) • FP = ([Gain, Normal, LOH, HOLoss],[AMP]) • TN = ([Gain],[GAIN, NEUT, HETD, HOMD]) + ([NEUT], [GAIN, NEUT, HETD, HOMD]) + ([LOH], [GAIN, NEUT, HETD, HOMD] + ([HOLoss], [GAIN, NEUT, HETD, HOMD]) 4. Homozygous deletions: • TP = ([HOLoss],[HOMD]) • FN = ([HOLoss],[AMP, GAIN, NEUT, HETD]) • FP = ([Amplification, Gain, Normal, LOH],[HOMD]) • TN = ([Amplifcation], [AMP, GAIN, NEUT, HETD]) + ([Gain],[AMP, GAIN, NEUT, HETD) + ([NEUT], [AMP, GAIN, NEUT, HETD]) + ([LOH], [AMP, GAIN, NEUT, HETD]) 5. Alterations: • TP = ([Amplification, Gain], [AMP, GAIN]) + ([LOH, HOLoss], [HETD, HOMD]) 15



doi:10.1038/nature10983

• FN = ([Amplification, Gain], [NEUT]) + ([LOH, HOLoss], [NEUT]) • FP = ([Normal, LOH, HOLoss], [GAIN, AMP]) + ([Amplification, Gain, Normal], [HETD, HOMD]) + ([Amplification, Gain], [HETD, HOMD]) + ([HOLoss, LOH], [GAIN, AMP]) • TN = ([Normal, NEUT]) 6. Gains: • TP = ([Amplification, Gain], [AMP, GAIN]) • FN = ([Amplification, Gain], [NEUT, HETD, HOMD]) • FP = ([Normal, LOH, HOLoss], [GAIN, AMP]) • TN = ([Normal, LOH, HOLoss], [NEUT, HETD, HOMD]) 7. Losses: • TP = ([LOH, HOLoss], [HETD, HOMD]) • FN = ([LOH, HOLoss], [AMP, GAIN, NEUT]) • FP = ([Amplification, Gain, Normal], [HETD, HOMD]) • TN = ([Amplification, Gain, Normal], [AMP, GAIN, NEUT]) Based on these indices and definitions, accuracy and precision values, as well as, sensitivity and specificity and the positive and negative predictive value, were estimated for CBS and HMM-derived somatic copy number calls (Figure S4a and Table S11). Additionally, we computed ROC curves for the CBS-derived CNAs using different thresholds to call alterations (Figure S4b).

11

The copy number landscape of breast cancer

11.1 The germline CNV landscape We interrogated the genomic architecture of primary tumours and normal samples to generate the first genome-wide breast cancer CNV map (Figure S6, Tables S12-S15). Since CNVs are generally thought to localise to non-genic regions, it is noteworthy that numerous cancer genes (including BCAS1, CCND2, EPHA3, ERBB4, ETV6, JAK1, JAK2, MET, PDGFRA, PML, PTEN, RET, TMPRSS2, WNK1) were each targeted in > 5% of samples (Figure S6a), usually resulting in gain or deletion of only a portion of the gene. Notably, the CNV landscape was found to be conserved across the intrinsic breast cancer subtypes (Figure S6b), whereas distinct somatic copy number profiles characterised the subgroups. A complete list of CNV frequencies, gene-centric Ensembl-based CNV mappings, and population distributions are presented in Tables S12-S15. Taken together with the observed moderate impact on expression variation noted in the eQTL studies, these findings hint that the common CNVs that can be identified on this platform are unlikely to account for a substantial portion of the missing heritability of breast cancer, as has been noted by others in the context of cancer and other complex diseases [31].

16



doi:10.1038/nature10983

11.2

The somatic CNA landscape

The minimal common regions of alterations [28] were computed according to ER status and for the PAM50 [11] expression-based intrinsic subtypes (Table S5), and recurrent alterations were summarized using a genecentric approach (Tables S4, S7). We then compared the underlying copy number profiles for each of the intrinsic subtypes (Figures S32, S33). Despite the observation that distinct copy number profiles characterised the subgroups, unsupervised clustering of genome-wide copy number data in the discovery cohort revealed additional heterogeneity within the subgroups (Figure S31). We also clustered the data on a combined measure of GI (Figure S26) to reveal nine major subgroups, three of which associated with poor outcome. In contrast to previous reports [32], these clusters did not recapitulate the intrinsic subtypes. For example, while the Basal-like tumours (n = 118) were relatively homogeneous, a low-GI subgroup was evident within them, as we noted previously [33]. ERBB2 copy number independent cases (n = 20) were also apparent within the PAM50 HER2-enriched subtype (n = 87). Examination of the Luminal tumours indicated a high number of cases with ERBB2 amplification (Luminal A, n = 54; Luminal B, n = 74), suggesting that refinement of the HER2-enriched intrinsic subtype is needed. Using refinements of CNA boundaries we enumerated the location and frequency of genes that may be disrupted by amplicon events. Certain genes were repeatedly broken, as a result of CNAs that occur within the gene such that a segment of copy number gain or loss is accompanied by a copy number neutral segment, disrupting its contiguous sequence. Recurrently broken genes include EGFR, ERBB2, FGFR1, IGF1R, PPP1R12B and PTPRT, whereas others (CCND1, ZNF703) were never broken and reflect amplicons that typically span many genes. The complete lists are enumerated in Tables S6 and S8.

12

eQTL analysis

Several eQTL analyses were performed using ANOVA, where the ranked expression values for each probe were regressed, in turn, on the set of CNV, CNA and SNP calls. The primary analysis relied on measurements derived from tumour tissue in the cohort of 997 individuals. However, as the individuals were of mixed ancestry, we also evaluated the potential influence of population structure on the results, by performing the analysis on an ethnically homogeneous population. Performing the analysis on a smaller cohort of European ancestry and for whom normal genotype data were available substantially reduced the sample size, and a decrease in the number of significant associations was anticipated. However, the relative contribution of different predictors was stable (Figure S12, Table S18), suggesting that the significance of the eQTLs was not inflated by population stratification. This analysis also lends support to our approach for dealing with the uncertainty of genotype calls derived from tumour tissue, wherein as a result of non-diploid genotype states in the tumours, SNP genotypes coincident with regions of copy number change might be altered. To address potential biases this might introduce, SNPs within 500 bp of a non-diploid CNV or CNA were set to missing. As this choice was heuristic, the suitability of this threshold was assessed by repeating the analysis on a set of 244 individuals of European ancestry and for whom ‘unbiased’ genotype calls had been obtained from normal breast tissue. As described below, this analysis indicates that only a minimal increase in precision could be expected had normal genotype data been available for all samples (Table S19). Finally, we investigated the robustness of the germline/somatic partitioning by considering a set of 85 samples, for which gene expression data were also available from normal breast tissue. Although the power for this analysis was markedly lower, it was assumed that strong germline associations present in the main analysis would persist, while associations due to somatic copy number changes would vanish, and indeed this was observed (Figure S13, Table S20).

17



doi:10.1038/nature10983

12.1

Preprocessing of predictors

Based on the annotation files employed, the SNP 6.0 chip provided measurements on 1,876,300 probes, including 906,600 polymorphic probes and 969,700 copy number, non-polymorphic probes. Since the CBSderived copy number calls provided the complete enumeration of germline CNVs and somatic CNAs for each sample, these events were analysed separately. Copy number states were collapsed into three levels, namely, loss (HETD, HOMD, or CNVLOSS), neutral (NEUT) or gain (GAIN, AMP, or CNVGAIN), coded as 0, 1, and 2, respectively. As noted above, it was possible for a region to contain both a CNV and CNA, typically due to a narrow CNV being flanked on either side by a CNA. Based on these data, CNV and CNA matrices were populated, both of size 997 × 1,876,300, where each row corresponds to a patient and each column a copy number probe, while the element records the copy number state (0, 1, or 2). Similarly, a SNP matrix of size 997 × 906,600 was constructed from the genotype calls derived from the SNP (polymorphic) probes on the chip, where each row corresponds to patient, each column a SNP probe, and each element records whether the genotype is homozygous (AA), heterozygous (AB), or homozygous (BB), coded as 0, 1, 2, respectively. Due to the fact that these genotypes derive from tumour tissue, we anticipate biases to arise when a SNP probe corresponds to a non-diploid copy number region, as a deletion/amplification will likely reduce/increase the allele count, respectively. For this reason, any individual’s SNP calls within 500 bp of such a region were set to missing, resulting in the loss of approximately 9.7% of genotypes. The second analysis using SNPs derived from non-cancerous tissue sought to investigate the extent of these biases. The logic for selecting a cut-off of 500 bp is noted below. The three (raw) predictor matrices were then condensed to obtain processed predictor matrices through a series of filtering steps. The first step removed trivial columns such as CNVs, CNAs, or SNPs that exhibited no variation across the 997 individuals. The second step removed duplicate columns. For copy number predictors, these typically corresponded to probes that lie within the same copy number region; for genotypes, these corresponded to SNPs in very high linkage disequilibrium (LD). Following these processing steps, 11,538 CNV, 193,873 CNA and 874,649 SNP predictors remained.

12.2

Preprocessing of responses

Given the large number of tests to be performed in the genome-wide setting, we sought to reduce this by eliminating potentially spurious expression probes. Following the strict criteria for exclusion of probes from the normalisation of the expression data, we also excluded probes that either contained a SNP, did not represent a perfect genomic match, were not 30 matching and without multiple genomic matches, contained a polyG tail at the end of probe sequence, or whose GC content was not between 38% and 64% (but retained those that targeted the sex chromosomes or were on the PAM50 genelist), resulting in 28,609 probes. The gene expression measurements were stored in a matrix of size 997 × 28,609, with again one row for each individual, and a column for each of the filtered responses. The eQTL analysis subsequently involved regressing each column of the response matrix against the columns of each predictor matrix. In order to identify the linear combinations of responses that exhibited the most variation across individuals, principal component analysis (PCA) was applied to the expression matrix. As samples derived from multiple sites, it was anticipated that this would be apparent in the top principal components, as was indeed the case. Additionally, ER status was expected to have a sizeable effect on expression. To account for these factors, the expression values were regressed on site and ER status using a categorical model (2 × 3 = 6 degrees of freedom). From this point on, the actual expression values were replaced by the corresponding residuals from this regression. Though it would have been possible to include site and ER status as cofactors in the subsequent association analysis, this approach reflects the belief that these two cofactors are likely to

18



doi:10.1038/nature10983

explain a given probe’s variation more than any single predictor and additionally it reduces the computation time significantly. Our original intention was to make use of the quantitative nature of gene expression measures, and regress the raw values directly on the predictors. However, we noted that the response matrix contained a large number of outliers that might be due to large changes in expression, but could also be attributed to measurement error. Such extreme values are highly undesirable when regressing on low-variance predictors. For example, suppose a predictor is observed in state 0 for only one individual (e.g. a copy number loss, which is observed only once). If a response were to have a single extreme value, then at random, the chance of this coinciding with the state 0 individual is 1 in 997, but when it does, basic regression methods will typically return a far more extreme P-value, usually leading to spurious results due to false associations. In these cases, generating P-values through permutation testing will often be more informative, as they will recognise the chance of such an occurrence [34]. However, when regressing each of many thousands of responses against many tens of thousands of predictors, meaningful permutation testing is not feasible. For this reason, we experimented with using raw values or introducing various corrections for outliers, and ultimately decided to convert the expression measurements to their corresponding ranked values such that the resulting analysis was essentially equivalent to performing a Kruskal-Wallis test, which is based on the one-way ANOVA formula using ranks [35]. We compared the ANOVA on ranks and the KW test and find that the latter produces, on average, slightly less extreme P-values. However, the conclusions from the eQTL analysis were nearly unchanged. The high degree of concordance between the P-values from ANOVA and those from a KW test, for CNV, CNA, and SNP associations is shown in Figure S11a. In all cases except a single SNP eQTL, the genes declared not to have a significant association by the ANOVA test were similarly not declared to have an eQTL association using KW. Relatively few (3/78, 180/10,863, 122/1,998) of the genes declared by ANOVA to have a significant CNV, CNA or SNP association were not significant in the KW test. Although there is a drop in concordance for extreme values (-log10 P-value >20), as with any asymptotic test, the accuracy of P-values deep into the tails is inevitably less reliable. We also compared the Venn diagrams showing the relative number of significant association for each predictor based on ANOVA (Figure 1) and the KW test (Figure S11b). While slightly fewer genes have significant eQTLs based on the KW test, the percentages in each cell stay almost constant. Below we explain how the use of ranks impacted the analysis.

12.3

eQTL analysis of tumour gene expression profiles

ANOVA was performed to regress each gene expression, in turn, on the three predictor sets (CNVs, CNAs, SNPs) (see Figure 1, Table S17) as described here for the case of CNVs. For each response, an ANOVA test was performed for each of the 11,538 CNV predictors, in turn. A likelihood ratio test was used to compare the null model of no association against a categorical alternative model. In particular, we assume that we have k random samples, one from each of k populations, where X j1 , . . . , X jn j is a random sample of size n j from the jth population, j = 1, . . . , k. We assume the jth population has median µ j and variance σ2 , and that the k random samples are independent. Here we note that under the null hypothesis that samples come from populations with identical locations, the distribution of the response is independent of the predictor state; while under the alternative hypothesis that at least one group differs in its distribution, its median varies according to which state is observed. Therefore if µ0 , µ1 , µ2 represent the median of each population group, we test H0: µ0 = µ1 = µ2 H1: µi ,µ j for some i , j The null model has degrees of freedom 2, one for the location and one for the variance. The alternative

19



doi:10.1038/nature10983

model, with degrees of freedom d+1 equal to the number of states observed for that predictor (typically d = 3) and one for the variance, allows each state to have an independent effect on the response. A test statistic was obtained as a function of the percentage of variance explained by the full model, which was compared to a χ2 distribution with degrees of freedom d-1 to obtain a P-value. For each response, the smallest of the 11,538 P-values was retained. Having performed this analysis for all 28,609 gene expressions, we assessed the evidence for significant CNV associations (likewise for CNAs and SNPs). Due to the large number of tests performed, it was necessary to correct for multiple testing. The Šidák correction [36] adjusts each P-value using a transformation of the form: p → 1-(1-p)Nr , where Nr represents the effective number of tests: N is the number of predictors, while r is a ratio between 0 and 1. The Šidák method is less conservative than the Bonferroni method, while guaranteeing strict control of the family-wise error rate (when comparisons are independent). If the tests for each of the 11,538 CNVs were independent, r = 1 would be suitable. However, as the tests were not independent, such a correction would be overly cautious, resulting in a bias of adjusted P-values close to 1 (Figure S10a). We attempted to find a suitable value of r using median mapping such that the median P-value is transformed to 0.5, similar to the genomic control approach [37]. This supposes that for most of the responses, the null hypothesis is true, in which case the corresponding P-value will be drawn from a uniform distribution on [0,1]. This transformation appears to perform reasonably well for the CNVs, where we anticipate that not many predictors will be associated, but again seems too conservative for CNAs and SNPs: if many of the P-values correspond to true alternative hypotheses, then the median of the correctly adjusted P-values will be significantly less than 0.5, so the r calculated from this method will be too large (Figure S10b). Q-Q plots of the adjusted P-values based on Šidák correction (r = 1) and median mapping were evaluated (Figure S10c). If P-values were correctly adjusted, then for values falling between quantiles 0.65-0.93, the QQ line should be straight in this region. The choice of quantiles, while fairly arbitrary, arose from considering that 8,000 P-values (0.28%), which should be more than sufficient to obtain a straight line, and should allow for a large enough window to judge a flat part of the distribution, an approach inspired by that used to calculate the q-value [38, 39]. We also sought to avoid P-values in the right tail, close to the upper boundary, in case the distribution was skewed for being too close to 1. We selected the value which maximised this correlation and so resulted in the straightest portion of line (Figure S10d). The values of r for CNVs, CNAs and SNPs (0.285, 0.045 and 0.67, respectively) correspond to the effective number of independent tests for each of these predictors (3,289, 8,724 and 586,014) and result in histograms with more desirable properties (Figure S10e).

12.4

Effect of ranking expression values

We performed a small study to test the effect of regression using ranked expression values as opposed to raw expression values. For a set of 300 expression probes, we concluded there was sufficient evidence for an association if the P-value was less than 0.001. For each probe, we then calculated an empirical P-value by regressing the raw expression values on the predictors and using 10,000 permutations. The permuted P-value was considered the gold standard, from which the true presence/absence of a CNV association was determined according to whether the empirical P-value was less than 0.001 or not (Figure S11c). When instead using P-values obtained by regressing ranked values on predictors, for 40 of the 300 probes (13%), a different conclusion was reached depending on the approach used. This figure should not be considered representative of an overall misclassification rate, as the 300 gene probes were selected on the basis that they were potentially troublesome (returning P-values close to the significance threshold). For the same set of 300 gene probes, the analysis was repeated using the raw expression values. This time, contradictory conclusions were reached for 173 of the 300 probes (58%), suggesting that outliers were common-place and

20



doi:10.1038/nature10983

that regression with ranked values was more reliable.

12.5

Definition of cis/trans windows

In the context of these association analyses, the definition of proximity is somewhat arbitrary, and the proportion of associations in cis will change as this window is varied. As the definition of the cis window is increased from 3 Mb (as reported for the main analysis), to 10 Mb, to the same chromosome, the number of trans associations will decrease, while the number of genome-wide associations remains constant (Figure S9a). An alternative approach would have been to perform the analysis twice: first considering only proximal predictors and second considering only distal predictors (Figure S9b). The first pass will be similar in spirit to the approach used by Stranger et al. [34], and would have greater power to detect cis associations by considering them on their own, rather than with all other predictors. However, since two eQTL analyses are being conducted for each predictor type, and potentially two predictors can be declared associated for each expression probe, it seemed prudent to tighten the declaration threshold to 0.00005, (i.e. half the previous cut-off). As expected, the values here are larger than when searching genome-wide (e.g. performing a single analysis and then partitioning associations into proximal and distal according to window size, as described above). However, they decrease as the definition of the proximal window is increased, showing that the increase in penalty for multiple testing is only partly offset by the extra signal included when widening the search. As the multiple testing penalty is nearly twice that used in the original analysis, it might be surprising that more associations are found for the distal associations as the proximal window is widened. A clue as to why this might be the case comes from considering the effect of increasing the definition of proximal. Here we observed that ‘distal predictors’ close to the proximal boundary are over-represented among those declared associated within the trans-associated group. Thus if the definition of proximal as events that occur on the same chromosome is considered overly cautious, this suggests there some trans signals will be obscured in the main analysis, presumably because they are overcome by more dominant cis signals.

12.6

Assessment of genotype calls in non-diploid regions and the impact of population stratification

Another analysis considered a set of 244 individuals of European-apparent ancestry and for whom SNP genotyping had also been performed on normal breast tissue (Figure S12, Table S18). As mentioned above, an element of doubt in the veracity of the genotype calls was introduced when a SNP coincided with a nondiploid (non-neutral) region of copy number variation. In the main analysis, these calls, and those within 500 bp of the region’s boundaries, were set to ‘missing’. The threshold of 500 bp was selected heuristically, by comparing the accuracy of SNP calls close to regions of copy number change. For the 244 individuals, the accuracy of the genotyping performed on tumour DNA was measured by assuming the calls obtained from non-tumour tissue were correct. The baseline miscalling rate calculated from all calls at least 50 kb away from the nearest copy number change was estimated to be 2.39%. The accuracy of calls with varying proximity to a non-diploid copy number region (CNV or CNA) was also computed (Table S19). For example, for calls lying within a region of copy number change (threshold 0), the miscalling rate was 8.13% for CNVs and 9.03% for CNAs, and these values diminished as the threshold was increased. These two figures are not independent, as SNPs can be near both a CNV and a CNA. To judge the best threshold cutoff, we considered the miscalling rate between consecutive thresholds. Ideally we would stop when this matched the baseline, however, as the decay of this rate to baseline seemed quite gradual for high thresholds, we selected 500 bp. Although at this distance the rate appears at least 1 point

21



doi:10.1038/nature10983

above baseline, we considered this a fair trade-off between accuracy and introducing too many missing values. The subset of 244 individuals was analysed via ANOVA in an identical fashion to the first analysis of 997 individuals. Although the detection power would be decreased, the expectation was that the ratio of gene probes found to have CNV, CNA, and SNP associations would remain similar, suggesting that the issue of dubious SNP genotyping calls in the presence of copy number variation was dealt with appropriately, and that population stratification did not inflate the significance of the eQTL results. When compared to the main analysis on 997 individuals, the number of gene probes with one or more association fell from 11,198 to 2,111. However, the relative contributions of different predictor types remain reasonably stable. Of the 2,111 gene probes, 90.5% have a CNA association, 9.8% have a SNP association and 0.5% have a CNV association (compared to 97%, 14.7% and 0.7%). However, the fact that the proportion has risen for CNAs while dropping for SNPs might suggest that with only 244 individuals, the chance of detecting two signals for a particular probe is greatly reduced and this is demonstrated by the sharp reduction in the proportion of gene probes with both a CNA and SNP association recorded (11.5% to 0.3%). Of the 208 gene probes found to have a SNP association, 204 were reported in the first analysis. That four were not reported reflects the fact that at a false positive rate (FPR) of 0.0001, we would expect 2.9 false calls from each analysis. Further examination of these results revealed that 3 of these 4 associations only just exceeded the declaration threshold (P-values between 0.00004 and 0.0001). As is evident from the variance histograms (Figure S12b), many of the top associations from the analysis of the 997 individuals (Figure 1) have persisted. These plots also highlight the effect of reducing the sample size. For example, for CNVs and CNAs the proportion of variance a predictor must explain to register as an association rises from approximately 3% to 12%, shown by movement of the left boundary of the CNA histograms. For SNPs, the cut-off rises from about 4% to 16%. These results lend support to our approach for dealing with the uncertainty surrounding genotype calls for SNPs in proximity to regions of copy number change, suggesting that the results from the main analysis are sensible and only a minimal increase in precision could be expected had normal genotypes have been available for all samples. Additionally, since the ethnic composition of the cohort was mixed, by restricting this analysis to samples of European ancestry (and for which normal genotype calls were available), we indirectly assessed whether population stratification inflated the significance of the eQTLs. Due to the diminished sample size, we anticipated that the number of significant associations would drop substantially (as a given predictor would be required to explain a greater proportion of the variance). However, as noted above, the relative contributions of each predictor remains stable. Since the accuracy of SNP calls, and hence SNP associations would be predominantly affected by non-diploid copy number regions, whereas CNA and CNV associations would be relatively unaffected, these results also suggest that population stratification has not significantly biased the results.

12.7

eQTL analysis of normal gene expression profiles

An additional analysis considered a set of 85 individuals of European ancestry for whom genotyping was performed on non-cancerous tissue and gene expression values from matched normal tissue were available (Figure S13, Table S20). The expectation was that strong germline associations would persist in this reduced sample set, while somatic associations would be ablated. As anticipated, the detection power was massively reduced, with an expression association declared for only 6 CNVs and 73 SNPs (compared to 78 and 1,617 from the main analysis). Of the 6 genes with a CNV association, 2 were identified in the main analysis, whereas 63 of the 73 SNP associations were also found by the main analysis, suggesting reasonably high concordance. As illustrated in the variance histograms (Figure S13b), the strongest CNV associated gene,

22



doi:10.1038/nature10983

GSTM1, is again identified, while IPO8 also persists in having the strongest SNP association. Reassuringly, expression probes found to have a CNA association have all but disappeared. Of the 5 that remain, 2 were also reported in the first analysis, suggesting only a small element of misclassification of somatic and germline events; the remaining 3 were expected when 28,609 tests are performed with a significance threshold of 0.0001.

12.8

Trans-acting aberration hotspots

Several trans-acting aberration hotspots that modulated the expression of many mRNAs were identified through the eQTL analyses, including the TCR-deletion hotspots on chromosomes 7 and 14 (Figure 3). As an aside, we note that the chromosome 14 hotspot was also detected when a novel approach to detect sparse regulatory networks [40] was applied to these data. Here, a cutoff of 30 was selected such that CNAs associated with > 30 mRNAs were termed a ‘hotspot’, resulting in 36 regions for both the 3 Mb window and 10 Mb window, respectively (33 of which overlap) (Table S25). We then performed enrichment analysis of the trans-associated mRNAs for the 10 Mb window size in order to identify pathways that might be coordinately regulated by a particular CNA hotspot (Table S26). Here, cognate mRNAs with a P-value < 10−4 , were taken to be significant for a given hotspot and were included in the analysis. Enrichment was assessed using the GOstats [41] package, where the set of 28,609 genes were used as the universe. Additionally, enrichment maps corresponding to modules enriched amongst trans associated mRNAs were generated with cytoscape. Briefly, genes were first projected onto a network using associations encoded in the Reactome FI Cytoscape plugin. The network was then clustered into modules using an edge-betweenness algorithm [42] maximising within module connectivity, while minimising between module connectivity, where pathways enriched (false discovery rate (FDR) < 0.001) for genes clustered in each module are shown. For each pathway, the source database is indicated as follows: B: NCI-PID BioCarta, C: cancer cell map, K: KEGG, N: NCI-PID curated pathways, P: PANTHER, R: Reactome.

13

Analysis of somatic copy number aberrations and gene expression

Motivated by our observation that proximal CNAs are a dominant feature of the tumour expression architecture (Figure 1, Figure 3a), we examined this trend in further detail. In particular, we collapsed the genome to a gene-centric scaffold and performed ANOVA, Kruskal-Wallis, and Spearman correlation tests of association between copy number and expression at the same gene (in cis) (Figures S29, S30, Tables S30-S37) and χ2 tests to determine subtype-specific CNAs within the integrative subgroups (Tables S38, S39). We also examined genome-wide patterns of copy-number and expression correlation (Figure S21). As was observed when the proximal window was set at 3 Mb, many of the top associations localise to genes on chromosomes 1, 8, 11, 16, and 17. Several examples of the relationship between CNAs and expression are illustrated in Figure S14. In particular, subtype-specific CNA-expression correlations are highlighted for ERBB2 and RSF1, which are amplified in HER2/Luminal and Luminal tumours, respectively. The expression of MAP2K4, which is deleted in Luminal cases, also shows copy number dependence, whereas TFF1 exhibits expression differences amongst the subtypes independent of copy number.

13.1

Subtype-specific analysis of copy number

A χ2 (Chi-Squared) test of independence was used to determine subtype association with copy number states derived from CBS or HMM. For each gene g, copy number values were divided into 3 levels: loss (HOMD & HETD), gain (GAIN & AMP/HLAMP) and neutral (NEUT). For the omnibus test, patients were divided 23



doi:10.1038/nature10983

into n (number of subtypes) levels; in the case of PAM50 subtypes, there are 5 (Basal, HER2, LumA, LumB, Normal) levels. For the subtype-specific test, patients were divided into 2 levels, subtype-of-interest or not. For these analyses, we employed the list of genes on the Illumina HT-12 v3 array that overlap with protein-coding genes from Ensembl 54 (hg18) (18,733; 18,670 excluding Y chromosome probes). Genes with insufficient representation in at least 2 copy number groupings (with a threshold of 10 cases) to perform the χ2 test (in the case of 3 groups) or Wilcoxon test (in the case of 2 groups) were excluded.

13.2

Cis-acting copy number and differential expression associations between groups

We collapsed the genome to a gene-based scaffold and investigated whether genes are differentially expressed in the tumours when samples are grouped based on predicted copy number states (derived from either CBS or HMM) as captured in the call matrix, C. To this end, we used several hypothesis-testing methods, namely ANOVA/Tukey-HSD and pairwise Student’s t tests (parametric) and Kruskal-Wallis/Wilcoxon rank sum tests (non-parametric). For each gene, samples were grouped into 3 levels based on their copy number alteration call: loss (HOMD & HETD), gain (GAIN & AMP/HLAMP) and neutral (NEUT). Then for each gene g and each probe p ∈ Pg , tests of differential expression were performed using Kruskal-Wallis and ANOVA between the groups defining the null hypothesis as drawing samples from the same population or as having the same rank means between the 3 groups, respectively. Often, genes were not altered with sufficient frequency (< 1%) in one level, hence reducing the problem to a Student’s t and Wilcoxon rank sum test for 2 groups. If the combined frequencies for both gains and losses were below 1%, then no hypothesis test was performed. Multiple testing correction was applied to all P-values using the Bonferroni method. For these analyses, we report the set of features given by Ensembl that correspond to Illumina HT-12 v3 probes (30,566) or those that map uniquely (29,033).

13.3

Cis copy number and expression correlation analysis

We also investigated the effect of copy number alterations on gene expression using Spearman’s rank correlation. Because the full enumeration of somatic and germline events is available, genes altered by CNAs and CNVs were analysed separately for both CBS and HMM. Spearman’s rank correlation tests were performed on each gene using the median log-ratios in the log matrix L and log2 expression values across samples. Because each gene g on the Illumina HT-12 v3 array may contain multiple probes Pg = {p1 , . . ., pi }, we assign gene g with the correlation coefficient, ρ (rho), and P-value from the probe p0 ∈ Pg with minimum P-value, (13) p0 = argmin{pvalue(p)} p∈Pg

There may be multiple p0 in the case of ties; thus, we select the probe p00 = argmax{rho(p)}

(14)

p∈p0

with max ρ to break the tie. P-values were corrected for multiple-testing using the Bonferroni adjustment. The list of 18,670 genes on the Illumina HT-12 v3 array that overlap with protein-coding genes from Ensembl 54 (hg18) was employed in this analysis.

13.4

Genome-wide copy number and expression correlation analysis

We also investigated genome-wide patterns of correlation between copy number aberrations and gene expression. Given expression and copy number data for 18,733 genes across 997 patients, we computed the 24



doi:10.1038/nature10983

Pearson correlation between pairs of genes using the median log-ratios in log matrix L and log2 expression values across patients. CNVs were ignored in this analysis and a 18,733 × 18,733 matrix with each element representing the correlation between copy number aberrations and expression of two genes was generated. Correlation P-values were adjusted for multiple hypothesis testing using the Benjamini-Hochberg method [43]. Observations with P-values > 0.05 were excluded and the resulting correlation matrix was used for downstream analysis. The matrix of copy number and expression correlation profiles were plotted for both CBS and HMM-derived CNAs (Figure S21). Enrichment maps corresponding to modules enriched amongst trans associated mRNAs were generated with cytoscape. Briefly, genes were first projected onto a network using associations encoded in the Reactome FI Cytoscape plugin. As in the eQTL analyses, the network was then clustered into modules using an edge-betweenness algorithm maximising within module connectivity, while minimising between module connectivity, where pathways enriched (FDR < 0.001) for genes clustered in each module are shown. For each pathway, the source database is indicated as follows: B: NCI-PID BioCarta, C: cancer cell map, K: KEGG, N: NCI-PID curated pathways, P: PANTHER, R: Reactome.

14 14.1

Cluster analysis Integrative clustering

Given that additional heterogeneity was evident within the intrinsic subgroups (Figure S31), and that the analysis of paired copy number and gene expression data suggested subtype-specific CNA-driven expression changes in cis, we sought to investigate whether this relationship could be used to further delineate breast cancer subgroups. To this end, we employed an integrative clustering framework (iCluster) described by Shen et al. [44] and explored various feature selection paradigms. The integrative clustering approach is based on a joint latent variable model that flexibly handles the associations between multiple data types, and reduces the dataset dimensionality. In particular, the K-Means approach is framed as a Gaussian joint latent variable model, where the maximum likelihood estimate (MLE) of the latent tumour subtypes is obtained through the expectation-maximisation (EM) algorithm. Following EM convergence, the class indicator matrix is obtained by performing K-Means on the resultant estimates. By penalising the log-likelihood with an L1 -norm regularisation parameter, λ, feature selection is inherent to the process and a sparse solution to the clustering is obtained. Here, we focus on the the use of cis-acting genes that exhibit significant associations with CNAs (CBSderived) across the entire cohort of tumours as determined by gene-centric ANOVA, as this reflects our assumption that copy number aberrations influence phenotypes through concomitant changes in expression. In particular, the 1,000 most significant Bonferroni adjusted P-value probes from the set of 30,566 Illumina probes (Table S30) were employed. Both copy number segmented means and normalised gene expression values for the selected features were used as input to the clustering, after removing copy number probes that exhibited high correlation and poorly annotated probes (caution was exercised in removing probes annotated to contain SNPs as these probes can interrogate key transcripts, and the frequency of the SNP is often sufficiently low in the population so not to impact the analysis). We examined differing numbers of clusters k ∈ [2, 18], where the sparsity parameter λ was determined by optimising the proportion of deviance (POD), defined by Shen et al. [44]. Selecting an appropriate number of features is somewhat arbitrary, so here we empirically assessed the adjusted Rand index (a measure of similarity between clusterings) for different integrative clustering runs with g = 500, 1,000, or 2,000 features as input. We observed that g = 1,000 allowed for a sufficient number of features without saturating the analysis, hence this feature set was employed in the subsequent analyses. 25



doi:10.1038/nature10983

As noted by Handl et al. [45] cluster analysis is a complicated and interactive process, and no entirely reliable method exists to identify the number of clusters in a dataset. Hence cluster analysis should always be performed for a sensible range of clusters as this facilitates an understanding of the operations of the algorithm and the identification of trends in the data. Furthermore, most internal cluster validity measures exhibit bias with respect to the underlying structure of the partitioning [45]. We evaluated the optimal number of clusters using Dunn’s index [46], which defines the ratio of the smallest distance between observations not in the same cluster to the largest intra-cluster distance, and thus assesses both intra-cluster homogeneity and inter-cluster separation. In particular, Dunn’s index was computed for the complete set of features used as input for integrative clustering using the Euclidean distance metric. Evaluation of the indices for k = 2, . . . , 15, indicated a global maximum at k = 14 (ignoring the k = 2 peak) and a local maximum at k = 10 (Figure S23a). The adjusted Rand index (ARI) was computed for all pairwise comparisons of k = 2, . . . , 18. Here we observe a relatively high adjusted Rand index for the k = 10 versus k = 9, 13, or 14 result, whereas that for k = 11 or 12 was lower (Figure S23b). We also inspected the similarities and differences in terms of cluster membership and genome-wide copy number profiles for alternate values of k, focusing in detail on k = 9, 13, 14 relative to the k = 10 solution. In particular, cross tabulation of the sample membership for the two clusterings indicate that the k = 9 partition exhibits a high degree of overlap with the k = 10 IntClust 1, 2, 4, 5, 6, 7, 8, and 9 (Figure S23c). However, IntClust 3 is split between several groups in the k = 9 clustering as is IntClust 10, which represented predominantly Basal-like tumours. Examination of the genome-wide copy number profiles associated with k = 9 suggest that they are less homogeneous than those observed for k = 10. A comparison of the k = 10 versus k = 13 result suggests that the 10 clusters (for the k = 10 solution) are generally recovered, and that the 3 new clusters labelled 6, 7, and 9 for k = 13 are not clearly distinct. For example, subgroup 6 is a mixture of IntClust 4 and IntClust 10 in the k = 10 grouping, and is similar to subgroup 5 for the k = 13 result. Similarly, subgroup 7 (n = 34) is composed primarily of IntClust 4 members, whereas group 9 (n = 56) is composed of IntClust 3 and IntClust 8, but exhibits a similar copy number profile to subgroup 3. Thus, k = 10 appears to capture the major features of the data observed for k = 13. A comparison of k = 10 and k = 14 yielded similar findings. Again, while the 10 clusters (for the k = 10 solution) are generally recovered, the 4 new clusters labelled 3, 5, 6, and 14 are not clearly distinct. Here, subgroup 3 (n = 87) is a mixture of IntClust 4 and IntClust 10 in the k = 10 grouping and is quite similar to subgroup 11. Subgroup 5 (n = 41) is a mixture of IntClust 6 and IntClust 9, whereas subgroup 6 is a mixture of IntClust 8 and IntClust9 (n = 60), and subgroup 14 is composed of multiple IntClust groupings, but primarily IntClust 3 and IntClust 8 and has a similar profile to subgroup 13. Thus, k = 10 also appears to capture the major features of the k = 14 clustering, and represents a conservative solution. A similar analysis was repeated for the HMM-derived CNAs (Figure S24). While many of the features of the CBS-derived integrative clustering results are also observed in this analysis, we note the absence of the 11q13/14 cis-acting subgroup for this partition. Thus, while alternative clusterings exist, we have presented one possible representation of the data that captures many salient features, providing insight into breast cancer biology. Perhaps most importantly, the clusters identified were found to reproduce in an external dataset as described below. In order to evaluate the reproducibility of the clustering results on an external dataset, we trained a classifier based on the 10 integrative cluster subgroups identified in the original dataset of 997 patients using the nearest shrunken centroids method described in Tibshirani et al. [47] and implemented in the Prediction Analysis of Microarrays (PAMR) software. Given that the integrative clustering approach has an inherent feature selection step, we used the set of 754 features selected by this method (where expression values were scaled) (Tables S42, S43) to build the PAM classifier, as any reduction significantly increased the error rate of the algorithm. Using a threshold value of 0.582 for shrinkage, the cross-validation error with 44 folds (as the smallest class size was n = 44) yielded a value of 0.087. Based on this set of centroids, we then classified

26



doi:10.1038/nature10983

the validation set of 995 tumours into the 10 groups (Table S44) and compared several attributes, such as the proportion of samples belonging to each group, their genome wide copy number profiles, Kaplan-Meier survival curves, and Cox models stratified by site with the following variables: grade, tumour size (modelled with a spline), number of lymph nodes (modelled with a spline), age at diagnosis (modelled with a spline), and integrative cluster membership (Figures 4, 5). Similar molecular characteristics were observed for each of the 10 clusters in both datasets, despite the fact that the validation set included low cellularity tumours. Moreover, Cox regression analysis shows similar fits for both datasets and the 95% confidence interval of the hazard ratios overlap (Table S40). The identification of reproducible clusters in an external dataset is related to prediction accuracy, which can be defined as the proportion of data whose predicted classifications are identical to the true classifications. Based on this relationship, Kapp and Tibshirani [48] define a procedure for the validation of clusters in a validation dataset, and a measure of cluster quality termed the in-group proportion (IGP), where IGP refers to the proportion of samples in a group whose nearest neighbours are also in the same group. Validation of groups found in an external dataset is performed by comparison against a null distribution of IGPs as implemented in the clusterRepro package in R, to provide a measure of individual cluster significance (as the IGP depends on the proportion of individuals belonging to each cluster and cluster sizes vary, the P-value should be used to assess cluster quality). These results demonstrate the reproducibility of the 10 integrative clusters, as all are deemed to be high-quality, although we note that the P-value associated with IntClust 3 is borderline significant. Thus the integrative clusters discovered in the training set can be reproduced in a validation set of tumours. It is worth noting that de novo clustering on a separate dataset does not address cluster reproducibility, but rather asks what can also be discovered in another cohort. Such an approach cannot be disentangled from the clustering procedure. It is not an approach that is powered for the validation of smaller patient sub-groupings because the inherent variability of clustering will not reveal all features in all subgroups, and will likely reveal additional groups. It is also not used in clinical practice, as one typically would want to assign a new (single) patient to a subgroup as can be done using a classifier. For this reason, internal cross validation and external validation on a separate dataset are the accepted methods for demonstrating the reproducibility of subgroups discovered by unsupervised clustering. Despite the limitations of this approach, de novo clustering can be performed on the validation set employed in this study. In doing so, we again observe that 10 represents the optimal number of clusters, confirming the clustering parameters (Figure S27). We also examined their molecular profiles and outcomes (Figure S27, Table S33), and note that the majority of the clusters are identified and have similar molecular profiles, although two of the smaller clusters originally identified in the discovery set (IntClust 2 and 6) are not well represented, presumably because the clustering favours a minimal number of samples per subgroup. A comparison with the predicted validation set assignments indicates reasonable agreement with 71.55% concordance. Analogous to the cluster validation procedure we describe above, one can build a classifier based on the features derived from integrative clustering and the cluster memberships. Using a threshold value of 0 for shrinkage, cross-validation error of 0.085 was obtained. Applying PAMR to the validation set we classified the discovery set and obtain 72.4% concordance with the de novo cluster membership in this dataset. All but the two smallest (IntClust 2 and 6) integrative subgroups were represented (as expected), and found to be reproducible based on the IGP. Thus, despite the fact that this approach would not be performed in practice to demonstrate subtype reproducibility, it recapitulates all of the larger clusters identified in the discovery set and highlight the robustness of these clusters since the validation set includes low cellularity tumours, which may influence feature selection and the clustering. The reproducibility of these integrative subgroups, which are characterised by distinct outcomes, suggests that by integrating multiple genomic features it may be possible to derive more robust patient classifiers

27



doi:10.1038/nature10983

that overcome the limitations of expression-only approaches. However, the development of a predictive signature is beyond the scope of this study. There are important conceptual differences between unsupervised class discovery methods, such as that applied here, and supervised clinical outcome prediction models or multigene prognostic signatures. The fundamental difference is that the latter employs supervised class prediction methods that take into account the known clinical outcome of cases during development of the predictor. These tasks are typically the subject of separate studies since they address distinct biological questions and utilise the data in very different ways. Nonetheless, it is possible to compare the predictive value of multivariable Cox models that include the integrative subtypes with the standard clinico-pathological variables using the concordance index (C-index) [49, 50]. Briefly, the multivariable Cox proportional hazards models were as follows (see also Survival analysis below): the basic model (m.basic) with clinical variables included the following components: Grade (numerical, linear), Size (numerical, spline), Lymph nodes (numerical, spline), and Age (numerical, spline), stratified by site; the PAM50 intrinsic subtype model is m.basic + Pam50Subtype, and the integrative cluster model is m.basic + IntClust. Cox models based on disease-specific survival were built for all cases as well as separately for ER-positive and ER-negative cases (as the proportional hazards assumption does not hold for ER status). The C-index computed for the discovery set indicates that the IntClust model has greater predictive power than the standard clinico-pathological variables alone and than the PAM50 model (Table S41a). For the validation set, the IntClust model offers improved predictive power over the standard clinico-pathological variables, and performs comparably to the PAM50 model (Table S41b).

14.2

K-Means clustering

We designed various feature sets for the K-Means clustering and examined different number of clusters k ∈ [5, 15]. Essentially, we employed gene expression profiles as features in the K-Means algorithm, and investigated different methods for identifying the most informative genes: • Entropy. Select those genes for which the entropy of the copy number profile is high, i.e. above half of the entropy of a uniform distribution. • Outlying. Select those genes which have been identified as outliers for extreme copy number states, based on the Detection of outlier expression profiles described below. We filtered the outlying genes P g∈Ok #(g,k) where Ok is the set of outlier genes based on the threshold of average count per gene τ = |Ok | for the copy number state k, and #(g, k) is the frequency of the outliers for gene g in the copy number state k. • cis-acting. Select cis-acting genes which exhibit high (absolute) correlation between copy number and expression, where the correlation is computed as described in the section on Correlation analysis. Here we employed the top 2,000 genes with the highest absolute Spearman correlation value. • Kurtosis. Select the top 2,000 genes with the least kurtosis. Comparison of these methods will be the subject of another contribution. For this paper, we report the results of the outlying features (Figure S25, Table S32). We note that while this approach captured some of the structure of the population, several key clusters identified based on integrative clustering were not observed.

28



doi:10.1038/nature10983

15

Detection of outlier expression profiles

To identify genes whose expression was driven in cis by extreme copy number events, Gaussian distributions were fitted to the 997 log2 expression values for each gene across the cohort, using the maximum likelihood estimates of the mean and variance. For genes with multiple probes, we summarized the expression level using the median. This resulted in 18,733 Gaussian distributions, each corresponding to the expression profile of a gene. We then sought to capture (gene, sample) events where large amplitude copy number aberrations (such as putative homozygous deletions and high-level amplifications) drive values into the extreme tails of the expression profile. Outlier events in the 5% right tail whose copy number state was either amplification or high-level amplification (AMP & HLAMP) as determined by HMM-Dosage and those events in the 5% left tail whose copy number state was homozygous deletion (HOMD) were reported along with the frequency of copy number aberration at that gene as determined by HMM and CBS (Tables S23, S24). We also illustrate the patterns of CNA-driven outlying expression at select regions throughout the genome in Figures S15 and S16.

16

Pathway analysis

Multiple tools and databases were employed to project genes that were significantly differentially expressed in a particular tumour subgroup relative to normal breast tissue (see the section on Gene expression analysis) onto pathways. Enrichment for Gene Ontology groups was performed using the NIH DAVID Bioinformatics resource (v6.7) [51, 52] and visualised with the Cytoscape software (v2.6.3) [53]. Pathway analysis was performed with Ingenuity Pathway Analysis software (v8.7) (Ingenuity Systems, www.ingenuity.com). The MetaCore (GeneGO) software, KEGG Pathways [54] and Wikipathways [55] databases were also used to cross-check the pathway components. Genes with a log2 ratio > 0.58 and an FDR adjusted P-value < 0.01, and that were associated with a canonical pathway in the Ingenuity Knowledge Base of canonical pathways and curated gene sets, were identified. The significance of the association between the data set and a canonical pathway was determined by computing the ratio of the number of features from the data set that map to the pathway divided by the total number of features that map to the canonical pathway. The probability that the association between the genes in the dataset and the canonical pathway could be explained by chance alone was assessed using Fisher’s exact test. The results of differential expression analysis and pathway enrichment in the integrative subgroups are presented in Tables S46 and S47, respectively. The relationships between molecules were represented graphically as a network, where molecules are represented as nodes, and the biological relationship between two nodes is represented as an edge (line). All edges are supported by at least one reference from the literature or canonical information stored in the Ingenuity Pathways Knowledge Base. Human, mouse, and rat orthologues are stored as separate objects in the Ingenuity Pathways Knowledge Base, but are represented as a single node in the network. The intensity of the node colour indicates the degree of up (red) or down (blue) regulation for a particular contrast. Nodes were displayed using various shapes that represent the functional class of the gene product. As described in the text, several subtype-specific gene networks were evident from pathway analysis of the CNA-expression landscape (Figures S36-S39). Note that for pathways that exhibited a major trans-acting component as determined by either the eQTL or correlation analyses, the relevant molecules are indicated with a green outline (Figures S36, S38).

29



doi:10.1038/nature10983

17

Survival analysis

We fit a series of Cox regression models for both disease-specific and overall survival using the survival [56] and Design [49,50] packages in R with several different perspectives in mind. First, we fit a global model for breast cancer specific survival for all 997 samples in the discovery cohort (for which 17 were removed due to missing values as they were lost to follow-up), including the most relevant clinical variables as covariates (grade, size, age at diagnosis, number of lymph nodes positive, ER status, and integrative cluster subtype). Models were also fitted to the validation cohort of 995 cases to evaluate cluster reproducibility in the context of clinical outcome (here 3 cases were lost to follow-up). Other variables were also missing for some cases. Analyses were stratified by tissue bank (site) to account for differences in the basal hazard due to geographical locations. We included spline terms for more complex numerical variables (tumour size, age at diagnosis, and number of lymph nodes positive). The residual analysis showed that ER status also violated the proportional hazards assumption, but the effect of this variable did not seem to be relevant and it was retained. For the analysis of genomic instability (GI), the effects for the GI index (jump) were added both globally and individually for each of the chromosome arms, keeping the significant ones (P-value < 0.05) for the multivariate model. Likelihood ratio tests were used to compare the models of clinical subtype, and combined clinical and molecular variables. Confidence intervals are reported for each of the variables, and differences in the survival distributions were also assessed using Kaplan-Meier curves and the nonparametric logrank test implemented in the survival package in R.

References [1] Zhou, W. et al. Full sequencing of TP53 identifies identical mutations within in situ and invasive components in breast cancer suggesting clonal evolution. Mol Oncol 3, 214–219 (2009). [2] Fridlyand, J. et al. Breast tumor copy number aberration phenotypes and genomic instability. BMC Cancer 6, 96 (2006). [3] R Development Core Team. R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria (2009). [4] Dunning, M. J., Smith, M. L., Ritchie, M. E. & Tavaré, S. beadarray: R classes and methods for Illumina bead-based data. Bioinformatics 23, 2183–2184 (2007). [5] Cairns, J. M., Dunning, M. J., Ritchie, M. E., Russell, R. & Lynch, A. G. BASH: a tool for managing BeadArray spatial artefacts. Bioinformatics 24, 2921–2922 (2008). [6] Asare, A. L., Gao, Z., Carey, V. J., Wang, R. & Seyfert-Margolis, V. Power enhancement via multivariate outlier testing with gene expression arrays. Bioinformatics 25, 48–53 (2009). [7] Barbosa-Morais, N. L. et al. A re-annotation pipeline for Illumina BeadArrays: improving the interpretation of gene expression data. Nucleic Acids Res 38, e17 (2010). [8] Smyth, G. K. Linear models and empirical bayes methods for assessing differential expression in microarray experiments. Stat Appl Genet Mol Biol 3, Article3 (2004). [9] Perou, C. M. et al. Molecular portraits of human breast tumours. Nature 406, 747–752 (2000).

30



doi:10.1038/nature10983

[10] Sorlie, T. et al. Gene expression patterns of breast carcinomas distinguish tumor subclasses with clinical implications. Proc Natl Acad Sci U S A 98, 10869–10874 (2001). [11] Parker, J. S. et al. Supervised risk predictor of breast cancer based on intrinsic subtypes. J Clin Oncol 27, 1160–1167 (2009). [12] Dunning, M. J. et al. The importance of platform annotation in interpreting microarray data. Lancet Oncol 11, 717 (2010). [13] Sørlie, T. et al. The importance of gene-centring microarray data. Lancet Oncol 11, 719–720 (2010). [14] Haibe-Kains, B. et al. A fuzzy gene expression-based computational approach improves breast cancer prognostication. Genome Biol 11, R18 (2010). [15] Haibe-Kains, B., Bontempi, G. & Sotiriou, C. genefu: Relevant Functions for Gene Expression Analysis, Especially in Breast Cancer. (2010). [16] Fraley, C. & Raftery, A. E. Model-based clustering, discriminant analysis and density estimation. Journal of the American Statistical Association 97, 611–631 (2002). [17] Fraley, C. & Raftery, A. E. MCLUST version 3 for R: Normal mixture modeling and model-based clustering. Technical Report 504, University of Washington, Department of Statistics (2006). [18] Lehmann, B. D. et al. Identification of human triple-negative breast cancer subtypes and preclinical models for selection of targeted therapies. J Clin Invest 121, 2750–2767 (2011). [19] Carvalho, B., Bengtsson, H., Speed, T. P. & Irizarry, R. A. Exploration, normalization, and genotype calls of high-density oligonucleotide SNP array data. Biostatistics 8, 485–499 (2007). [20] Irizarry, R. A., Carvalho, B. S., Scharpf, R. & Ritchie, M. crlmm: Genotype Calling (CRLMM) and Copy Number Analysis tool for Affymetrix SNP 5.0 and 6.0 and Illumina arrays. (2010). [21] Venkatraman, E. S. & Olshen, A. B. A faster circular binary segmentation algorithm for the analysis of array CGH data. Bioinformatics 23, 657–663 (2007). [22] Shah, S. P. et al. Integrating copy number polymorphisms into array CGH analysis using a robust HMM. Bioinformatics 22, e431–9 (2006). [23] Bengtsson, H., Wirapati, P. & Speed, T. P. A single-array preprocessing method for estimating fullresolution raw copy numbers from all Affymetrix genotyping arrays including GenomeWideSNP 5 & 6. Bioinformatics 25, 2149–2156 (2009). [24] Smith, M. L., Marioni, J. C., McKinney, S., Hardcastle, T. & Thorne, N. P. snapCGH: Segmentation, normalisation and processing of aCGH data. (2009). [25] Seshan, V. E. & Olshen, A. DNAcopy: DNA copy number data analysis (2010). [26] Ostrovnaya, I., Nanjangud, G. & Olshen, A. B. A classification model for distinguishing copy number variants from cancer-related alterations. BMC Bioinformatics 11, 297 (2010). [27] Willenbrock, H. & Fridlyand, J. A comparison study: applying segmentation to array CGH data for downstream analyses. Bioinformatics 21, 4084–4091 (2005). 31



doi:10.1038/nature10983

[28] Rouveirol, C. et al. Computation of recurrent minimal genomic alterations from array-CGH data. Bioinformatics 22, 849–856 (2006). [29] Conrad, D. F. et al. Origins and functional impact of copy number variation in the human genome. Nature 464, 704–712 (2010). [30] Bignell, G. R. et al. Signatures of mutation and selection in the cancer genome. Nature 463, 893–898 (2010). [31] Wellcome Trust Case Control Consortium et al. Genome-wide association study of CNVs in 16,000 cases of eight common diseases and 3,000 shared controls. Nature 464, 713–720 (2010). [32] Habermann, J. K. et al. The gene expression signature of genomic instability in breast cancer is an independent predictor of clinical outcome. Int J Cancer 124, 1552–1564 (2009). [33] Chin, S. F. et al. High-resolution aCGH and expression profiling identifies a novel genomic subtype of ER negative breast cancer. Genome Biol 8, R215 (2007). [34] Stranger, B. E. et al. Relative impact of nucleotide and copy number variation on gene expression phenotypes. Science 315, 848–853 (2007). [35] Kruskal, W. H. A nonparametric test for the several sample problem. The Annals of Mathematical Statistics 23, 525–540 (1952). [36] Šidák, Z. Rectangular confidence regions for the means of multivariate normal distributions. American Statistical Association 62, 626–633 (1967). [37] Devlin, B. & Roeder, K. Genomic control for association studies. Biometrics 55, 997–1004 (1999). [38] Storey, J. D. & Tibshirani, R. Statistical significance for genomewide studies. Proc Natl Acad Sci U S A 100, 9440–9445 (2003). [39] Storey, J. D. The positive false discovery rate: a Bayesian interpretation and the q-value. Annals of Statistics 32, 2013–2035 (2003). [40] Yuan, Y., Curtis, C., Caldas, C. & Markowetz, F. A sparse regulatory network of copy-number driven gene expression reveals putative breast cancer oncogenes. IEEE/ACM Trans Comput Biol Bioinform (2011). [41] Falcon, S. & Gentleman, R. Using GOstats to test gene lists for GO term association. Bioinformatics 23, 257–258 (2007). [42] Wu, G., Feng, X. & Stein, L. A human functional protein interaction network and its application to cancer data analysis. Genome Biol 11, R53 (2010). [43] Benjamini, Y. & Hochberg, Y. Controlling the false discovery rate: A practical and powerful approach to multiple testing. Journal of the Royal Statistical Society B, 289–300 (1995). [44] Shen, R., Olshen, A. B. & Ladanyi, M. Integrative clustering of multiple genomic data types using a joint latent variable model with application to breast and lung cancer subtype analysis. Bioinformatics 25, 2906–12 (2009).

32



doi:10.1038/nature10983

[45] Handl, J., Knowles, J. & Kell, D. B. Computational cluster validation in post-genomic data analysis. Bioinformatics 21, 3201–12 (2005). [46] Dunn, J. C. Well separated clusters and optimal fuzzy partitions. Journal of Cybernetics 4, 95–104 (1974). [47] Tibshirani, R., Hastie, T., Narasimhan, B. & Chu, G. Diagnosis of multiple cancer types by shrunken centroids of gene expression. Proc Natl Acad Sci U S A 99, 6567–6572 (2002). [48] Kapp, A. V. & Tibshirani, R. Are clusters found in one dataset present in another dataset? Biostatistics 8, 9–31 (2007). [49] Harrell, F. E. Regression modeling strategies: with applications to linear models, logistic regression, and survival analysis (Springer, New York, 2001). [50] Harrell, F. E. Design: Design Package (2009). [51] Huang, D. W., Sherman, B. T. & Lempicki, R. A. Systematic and integrative analysis of large gene lists using DAVID bioinformatics resources. Nature Protocols 4, 44–57 (2009). [52] Dennis, G., Jr et al. DAVID: Database for annotation, visualization, and integrated discovery. Genome Biol 4, P3 (2003). [53] Cline, M. S. et al. Integration of biological networks and gene expression data using Cytoscape. Nat Protoc 2, 2366–2382 (2007). [54] Kanehisa, M., Goto, S., Furumichi, M., Tanabe, M. & Hirakawa, M. KEGG for representation and analysis of molecular networks involving diseases and drugs. Nucleic Acids Res 38, D355–360 (2010). [55] Pico, A. R. et al. Wikipathways: pathway editing for the people. PLoS Biol 6, e184 (2008). [56] Therneau, T. & original R port by Thomas Lumley. survival: Survival analysis, including penalised likelihood. (2009).

33



doi:10.1038/nature10983

18

Supplementary Figures

34



doi:10.1038/nature10983

35



doi:10.1038/nature10983

Discovery cohort Primary raw data

Summarised data

Affymetrix SNP 6.0 raw data

Illumina HT-12 v3 raw data

Probe-level modeling, normalisation, summarisation of Log2 intensities with SNP-RMA / aroma.affymetrix

Spatial artifact correction, summarisation, normalisation of Log2 intensities with beadarray and bash

997 primary breast tumours 473 normals

997 primary breast tumours 144 normals

Copy number segmentation SNP genotype with CBS and HMM

Biological features

Validation

calling with crlmm

PAM50 classification

Partitioning into germline (CNV) & somatic (CNA) MLPA, PCR, events

Differential expression analysis

Sanger sequencing

Validation

Generation of a cancer CNV map Integrative analyses

qRT-PCR

eQTL analysis of SNPs, CNVs, CNAs

Impact on tumour expression architecture

Identification of cis-acting CNAs & outlier expression profiles

Identification of recurrent, high-level somatic aberrations Identification of Trans-acting aberration hotspots

Integrative clustering on cis-acting features

Assessment of cluster validity

Assimilation of clinical covariates & orthogonal data Validation cohort

995 primary breast tumours

Delineation of novel subgroups Mapping onto molecular pathways

Elucidation of subtype-specific cis/trans gene networks

Mapping onto molecular features

External validation of cluster reproducibility

Identification and refinement of subtype-specific drivers

The somatic landscape of breast cancer Figure S1: Workflow describing the main analytical approaches used in this study. 36



doi:10.1038/nature10983

37



doi:10.1038/nature10983

Figure S2: Assignment of cases to the PAM50 intrinsic subtypes. a. Heatmap illustration of gene expression values for the published PAM50 centroids. b. Unsupervised clustering of samples according to the PAM50 genes, where the dendrogram reflects the PAM50 assignments, grade, and ER status. c. Comparison of various assignments for key clinical markers based on immunohistochemistry (IHC), gene expression, PAM50, and copy number calls (SNP6) for ER status (upper panel), HER2 status (middle panel), and PR status (lower panel). 38



doi:10.1038/nature10983

39


doi:10.1038/nature10983


Figure S3: Verification of germline CNVs identified from the analysis of Affymetrix SNP 6.0 copy number data. a. Probe-level copy number plots (derived from CBS) illustrating several samples with SMARCA2 intronic deletion. b. Long-range PCR confirmed a 3,914 bp deletion. c. The deletion was mapped to nucleotide resolution by sequencing. d. Probe-level copy number plots (derived from CBS) illustrating a deletion between two olfactory receptor genes, OR51A4 and OR51A2 in a tumour and its matched normal. e. The deletion was captured using long-range PCR, and sequencing revealed a 8,585 bp deletion resulting in fusion of both olfactory receptor genes. 40



doi:10.1038/nature10983

−1

0

CBS

1

2

a

HOLoss

LOH

Normal

Gain

Amplification

Ambiguous

Gain

Amplification

Ambiguous

−1

0

HMM

1

2

MLPA

HOLoss

LOH

Normal MLPA

41



doi:10.1038/nature10983

1.0 0.4 0.2

0.2 0.4

0.6

0.8

1.0

0.0

0.0 0.2

0.0

0.2

0.4

1−Specificity

0.6

0.8

1.0

0.0

0.8

1.0

0.8

1.0

1.0 0.8

Sensitivity

0.8

0.4

0.6

Sensitivity

0.2

0.4

0.0

0.2 0.0 0.6

0.6

Losses+Homozygous deletions

1.0

1.0 0.8 0.6 0.4 0.2

0.4

0.4

1−Specificity

Gains+Amplifications

0.0

0.2

0.2

1−Specificity

Alterations

0.0

0.6

Sensitivity

0.8

0.8 0.4

0.6

Sensitivity

0.8 0.6

Sensitivity

0.4 0.2 0.0 0.0

Sensitivity

Homozygous deletions

1.0

Amplifications

1.0

HighLevel

0.6

b

0.0

0.2

0.4

1−Specificity

0.6

1−Specificity

0.8

1.0

0.0

0.2

0.4

0.6

0.8

1.0

1−Specificity

Figure S4: Assessment of somatic copy number segmentation for CBS and HMM. a. Comparison of CNA calling accuracy for CBS and HMM based on MLPA data. b. ROC curves for calling alterations based on CBS segmentation.

42



0.8

1.0

b

0.2

M W Basal

M W Her2

M W LumA

0.0

0.0

0.2

Proportion 0.4 0.6

Proportion 0.4 0.6

0.8

a

1.0

doi:10.1038/nature10983

M W M W LumB Normal

PAM50 group

MW MW MW MW MW MW MW MW MW MW 9 6 1 5 10 7 4 8 2 3

IntClust group

Figure S5: The proportion of TP53 mutated versus wildtype cases. a. Mutational frequencies across the PAM50 intrinsic subtypes. b. As in (a), but for the integrative cluster subgroups.

43



doi:10.1038/nature10983

Figure S6: The copy number landscape of breast cancer. a. Global map of copy number variation in breast cancer. Genome-wide probe-level frequencies of somatically acquired CNAs (upper panel) and inherited germline CNVs (lower panel) are indicated for gains (red) and losses (blue) (derived from CBS), where known cancer genes targeted by CNVs in > 5% of cases are labelled. b. Germline CNV landscape across the PAM50 intrinsic subtypes. Genome-wide probe-level frequencies of inherited germline CNVs, where gains are indicated in red and losses in blue.

44



doi:10.1038/nature10983

Asian

0

African

−50000

Second Principal Component

50000

Tumours

European −150000

−100000

−50000

0

50000

100000

First Principal Component

Basal

Her2

LumA

LumB

Normal−like

Figure S7: Primary HapMap population clusters for the METABRIC cohort.

Figure S8: Germline and somatic variants influence tumour expression architecture. Manhattan plots indicate the significance of genome-wide associations for each predictor type. The directionality of the associations is indicated as follows: cis - positive (red), negative (pink); trans- positive (blue), negative (green).

45



doi:10.1038/nature10983

a

46



doi:10.1038/nature10983

b

Figure S9: Venn diagrams based on alternative definitions of the proximal cis-window. a. Venn diagrams illustrate the relative contribution of germline variants (SNPs, CNVs) and somatic aberrations (CNAs) on genome-wide, cis, and trans tumour expression variation for three different proximal window sizes (3 Mb, 10 Mb, and same chromosome, rows 1-3, respectively) at a 0.0001 significance threshold after Šidák adjustment. This figure illustrates how the number of proximal associations increases, while the number of distal associations decreases, as the definition of proximal is enlarged. Note that for all three thresholds, the number of genome-wide association remains the same (column 1). An alternative approach would have been to perform the analysis twice: first considering only proximal predictors and second considering only distal predictors. Note that for this analysis the significance threshold was halved (reduced to 0.00005) to account for two separate tests being performed. b. The results of such an approach are illustrated here. Here, the three rows consider different definitions of proximal (3Mb, 10Mb and same chromosome). The first column reports the types of associations found when examining all predictors, the second column reports the types of associations found when examining only proximal predictors, and the third column reports the types of associations found when examining only distal predictors.

47



800 0.0

0.2

0.4

0.6

0.8

1.0

0.8

1.0

800 0.4

0.6

0.8

1.0

1.0

ca1a Plots CNA QQ

cs1[keep[ord]]

0.8 0.6

0.8

1.0

0.2

0.4

0.6

0.8

1.0

1.0

1:l/l vs QQ Fit CNA − Ratio

0.9

srec[, 2]

0.8

0.8

1.0

0.0

0.2

0.4

0.6

0.8

1.0

1:l/l vs QQ Fit SNP − Ratio

0.5

0.94

0.980

0.6

0.7

arec[, 2]

0.6

0.2 0.0

0.98 0.97 0.96 0.95

0.4

cs1aPlots SNP QQ

1.000

0.6

0.2

0.995

0.4

1:l/l vs QQ Fit CNV − Ratio

0.0

0.0

0.2 0.0 0.2

1.0

600

Frequency 0.2

0.4

ca1[keep[ord]]

0.6 0.4 0.2 0.0

d

0.99

1.00

0.0

0.8

200 0.0

1.0

0.6

0.6

0.8

0.4

cv1a Plots CNV QQ

0.4

cs1 Transform SNP − Median

0

200 0 0.2

0.2

400

800 600

Frequency

400

600 400 200 0

c

0.8

1.0

0.0

0.0 1000

ca1 Transform CNA − Median

0.6

1.0

0.4

0.8

0.990

0.6

800

1000

0.4

cv1 Transform CNV − Median

600

Frequency

200 0

0 0.2

b

1000

0.0

SNP − Sidak

400

800 600

Frequency

200

400 0

200

400

600

800

CNA − Sidak

0.985

CNV − Sidak

a

1000

1000

1000

doi:10.1038/nature10983

0.0

0.2

0.4

0.6

0.8

1.0

arec[, 1] Number CNA − Effective

800

Frequency

600

0.4

0.6

0.8

1.0

0.4

0.6

0.8

1.0

srec[, 1] Number SNP − Effective

0

200 0 0.2

0.2

200

Frequency

400

600 400 200 0 0.0

0.0 1000

1.0

800

0.8

600

0.6

800

1000

0.4

vrec[, 1] Number CNV − Effective

400

0.2

e

1000

0.0

0.0

0.2

0.4

0.6

0.8

1.0

0.0

0.2

0.4

0.6

0.8

1.0

48



doi:10.1038/nature10983

Figure S10: Assessment of various multiple testing adjustments for the eQTL analysis. a. Histograms illustrating the effect of the Šidák correction, which adjusts each P-value according to a transformation of the form: p → 1-(1-p)Nr , where Nr represents the effective number of tests, and r is a value between 0 and 1, shown here (r = 1). This appears to bias the adjusted P-values towards 1, and hence is overly cautious. b. Histograms illustrating attempts to estimate a suitable value of r via median mapping such that the median P-value is transformed to 0.5. While this approach seems suitable for the CNVs, where for many of the responses the null hypothesis is true, it is overly conservative for CNAs and SNPs. c. QQ plots of adjusted P-values, where the overly-cautious nature of setting r = 1 is indicated by the green line, whereas the blue line represents median mapping, which forces the Q-Q line to pass through the point (0.5,0.5). The adjusted P-values falling between the two dashed lines correspond to those ranked 18,610 to 26,609 (between quantiles 0.65 to 0.93). The assumption is that these 8,000 P-values likely correspond to true null hypotheses, so the line in this region should be straight. d. Q-Q plots illustrating a measure of ‘straightness’ for this region based on the R2 for different values of r. The value which maximised this correlation and so resulted in the straightest portion of line was selected. Notably, the values of r for CNVs, CNAs and SNPs were 0.285, 0.045 and 0.67, respectively and corresponds roughly to the effective number of independent tests (3,289, 8,724 and 586,014, respectively). e. This panel illustrates that the approach based on the effective number of tests resulted in histograms with more desirable properties for all three predictor types.

49



doi:10.1038/nature10983

a

11015 Genes with a Gwide Ass. CNVs

32 0.3%

4 0%

13 0.1%

26 0.2%

316 2.9%

CNAs 9458 85.9%

5883 Genes with a Cis Ass. CNVs 4 0.1%

1166 10.6%

12 0.2%

6 0.1%

12 0.2%

506 8.6%

SNPs

5773 Genes with a Trans Ass. CNAs

5029 85.5%

314 5.3%

CNVs 25 0.4%

7 0.1%

SNPs

5 0.1%

4 0.1%

420 7.3%

CNAs 5066 87.8%

246 4.3%

SNPs

b

50



doi:10.1038/nature10983

c

Ranked Expression − 29/300 Misclassified

3.5

0

2.5

3.0

−log10 Permutation Scores

3 2 1

−log10 Permutation Scores

4.0

Raw Expression − 175/300 Misclassified

20

40

60

80

100

120

6.5

−log10 ANOVA Scores

7.0

7.5

8.0

8.5

9.0

−log10 ANOVA Scores

Figure S11: Comparison of the results of ANOVA on ranked values versus the Kruskal-Wallis test and assessment of the use of ranked versus raw expression values for handling outlier responses in the eQTL analysis. a. Venn diagrams showing the relative number of significant association for each predictor based on the KW test. Note that although slightly fewer genes exhibit significant eQTLs relative to the ANOVA on ranked expression results, the percentages in each cell remain relatively constant. b. Scatterplot of -log10 P-values for the Kruskal-Wallis test versus ANOVA on ranked expression values indicates a high degree of concordance, although there is a slight drop for extreme -log10 P-values (> 20), as expected with any asymptotic test. The red lines indicate the significance threshold (0.0001 after multiple correction). c. Plots illustrating the difference between the sets of P-values obtained when ranked versus raw expression values are employed for the ANOVA analyses based on the computation of empirical P-values via permutation testing for a subset of 300 expression probes (see Methods). The permuted P-value was considered the ground truth, from which the presence/absence of an association was determined according to whether the empirical P-value exceeded 0.001 or not. Red points indicate a misclassification.

51



doi:10.1038/nature10983

2,111 Genes with a Gwide Ass.

9 0.4%

b

1901 90.1%

192 9.1%

1,042 Genes with a Cis Ass.

CNAs

CNVs

8 0.8%

7 0.3% SNPs

10 20 30

10 20 30

GSTT1

SNPs

40

60

80

0 0%

14 1.3%

SNPs

MAPK8IP1

0

0

ZC3H3

MAF1

TBCE

0

0

GPATCH4

60

WARS2

LOC650298 LOC651894 MAPK8IP1

20

RPL14

0

20 20

CNAs

SNPs − 15 Genes Trans Ass.

0

0 0

1057 98.6%

CNAs − 1057 Genes Trans Ass.

C1ORF131

VPS72

60 40

60 40

WARS2

0 0% 0 0%

CNVs − 1 Genes Trans Ass.

GSTM1

SNPs − 193 Genes Cis Ass.

20

GPATCH4

0 0% 1 0.1%

4 0.4%

GSTT1

MTERFD1

SNPs − 208 Genes Gwide Ass. RPL14

CNVs

CNAs − 851 Genes Cis Ass. 20 40 60

20 40 60

C1ORF131

VPS72

MTERFD1

0

Frequency

181 17.4%

HLA−DRB5

CNAs − 1908 Genes Gwide Ass.

Frequency

847 81.3%

1,072 Genes with a Trans Ass.

CNAs

CNVs − 10 Genes Cis Ass.

GSTM1

0

Frequency

CNVs − 11 Genes Gwide Ass. HLA−DRB5

2 0.2%

0 0% 0 0%

10 20 30

2 0.1%

0 0% 0 0%

20 40 60

CNVs

40

a

0

20

Variance

40 Variance

60

80

0

20

40

60

80

Variance

Figure S12: eQTL analysis in cases of European ancestry for which normal genotype calls were available (n = 244). a. Venn diagrams depict the relative contribution of germline variants (SNPs, CNVs) and somatic aberrations (CNAs) on genome-wide, cis, and trans tumour expression variation. Expression profiles that were significantly associated with either a CNV, CNA, or SNP at an adjusted P-value of 0.0001 are reported. b. Histograms illustrate the proportion of variance explained by the most significantly associated predictor for each predictor type, where several of the top associations are indicated. As above, the results are shown for the genome-wide analysis as well as for proximal (cis) and distal (trans) associations.

52



doi:10.1038/nature10983

82 Genes with a Gwide Ass. CNVs

4 4.9% 2 2.4%

b

0 0% 0 0%

CNAs

5 6.1%

71 86.6%

58 Genes with a Cis Ass. CNVs

2 3.4%

0 0% SNPs

SNPs

40

60

PHLDB1

HS.546213

20 40 60

PRH2

0

C15ORF56 HS.282795

GPATCH4

60

KRT82

LOC650298 MAPK8IP1

20

PTER

40

60 80

SNPs

0

0 0 20

0 0%

SNPs − 15 Genes Trans Ass.

IPO8

0

0

0 0

CNAs

CNAs − 5 Genes Trans Ass.

20

GPATCH4

20

PTER

5 20.8%

15 62.5%

DENR

SNPs − 58 Genes Cis Ass. 40

40

60

SNPs − 73 Genes Gwide Ass. IPO8

0 0% 0 0%

CNVs − 4 Genes Trans Ass.

20 40 60

20 40 60

Frequency

C15ORF56

4 16.7% 0 0%

CNAs − 0 Genes Cis Ass.

PRH2

0

HS.282795

CNVs

0 0%

GSTM1

CNAs − 5 Genes Gwide Ass.

Frequency

56 96.6%

CNAs

0 0%

10 20 30

10 20 30

HS.546213

0 0% 0 0%

CNVs − 2 Genes Cis Ass.

GSTM1

0

Frequency

CNVs − 6 Genes Gwide Ass. PHLDB1

0 0%

24 Genes with a Trans Ass.

10 20 30

a

0

20

Variance

40 Variance

60

80

0

20

40

60

80

Variance

Figure S13: eQTL analysis of normal expression data. This analysis was performed in cases of European ancestry for which normal genotype calls and normal gene expression profiles were available (n = 85). a. Venn diagrams depict the relative contribution of germline variants (SNPs, CNVs) and somatic aberrations (CNAs) on genome-wide, cis, and trans tumour expression variation. Expression profiles that were significantly associated with either a CNV, CNA, or SNP at an adjusted P-value of 0.0001 are reported. b. Histograms illustrate the proportion of variance explained by the most significantly associated predictor for each predictor type, where several of the top associations are indicated. As above, the results are shown for the genome-wide analysis as well as for proximal (cis) and distal (trans) associations.

53



doi:10.1038/nature10983

Figure S14: Illustration of cis-acting copy number expression associations. Scatterplots of scaled gene expression values versus segmented copy number values (log base 2, CBS-derived) for several exemplary genes, including ERBB2 and RSF1, which are amplified, MAP2K4, which is homozygously deleted, and TFF1, which shows copy number independent expression changes. Data points are colour-coded according to the PAM50 subtypes.

54


Genes & Chi−square Test

JAK1

RAVER2

CACHD1

UBE2U

ROR1

PGM1

DLEU2L

EFCAB7

ITGB3BP

ALG6

FOXD3

ATG4C

ANGPTL3

DOCK7

F

JAK1

RAVER2

CACHD1

UBE2U

ROR1

PGM1

DLEU2L

EFCAB7

ITGB3BP

ALG6

FOXD3

ATG4C

ANGPTL3

DOCK7

F

JAK1

RAVER2

CACHD1

UBE2U

ROR1

PGM1

DLEU2L

EFCAB7

ITGB3BP

ALG6

FOXD3

ATG4C

ANGPTL3

DOCK7

F

8

6

4

2

0

8

6

4

8

6

4

2 q44

q42.11 q42.12 q42.13 q42.2 q42.3 q43

q41

q32.2 q32.3

q32.1

q31.1 q31.2 q31.3

q25.3

q21.1 q21.2 q21.3 q22 q23.1 q23.2 q23.3 q24.1 q24.2 q24.3 q25.1 q25.2

q12

p13.3 p13.2 p13.1 p12 p11.2 p11.1 q11

p22.3 p22.2 p22.1 p21.3 p21.2 p21.1

p31.1

p31.2

p32.3 p32.2 p32.1 p31.3

p34.2 p34.1 p33

p36.33 p36.32 p36.31 p36.23 p36.22 p36.21 p36.13 p36.12 p36.11 p35.3 p35.2 p35.1 p34.3

Chromosome 1 amplification outliers

●

PAM50 region1 chr1:62692985−65465763 Subtypes Segregation − from DOCK7 to JAK1 PAM50 Normal LumB LumA Her2 Basal

2

0


IntClust region1 chr1:62692985−65465763 Subtypes Segregation − from DOCK7 to JAK1 IntClust 1 2 3 4 5 6 7 8 9 10

0


k−means region1 chr1:62692985−65465763 Subtypes Segregation − from DOCK7 to JAK1

k−means

9 8 7 6 5 4 3 2 1

10

S100A1 C1orf77 SNAPIN ILF2 NPR1 INTS3 SLC27A3 RP11−216N14.5 GATAD2B CRTC2 SLC39A1 CREB3L4 JTB RAB13 RPS27 NUP210L TPM3 C1orf189 C1orf43 UBAP2L HAX1 AQP10 ATP8B2 RP11−350G8.3 IL6R SHE TDRD10 UBE2Q1 CHRNB2 ADAR KCNN3 PMVK PBXIP1 PYGO2 SHC1 CKS1B FLAD1 LENEP ZBTB7B DCST2 DCST1 ADAM15 EFNA4 EFNA3 EFNA1 RAG1AP1 DPM3 KRTCAP2 TRIM46 MUC1 THBS3 MTX1 GBA C1orf2 SCAMP3 CLK2 HCN3 PKLR FDPS ASH1L MSTO1 YY1AP1 DAP3 GON4L SYT11 RIT1 KIAA0907 RXFP4 ARHGEF2 SSR2 UBQLN4 ROBLD3 RAB25 LMNA SEMA4A SLC25A44 PMF1 PAQR6 SMG5 C1orf85 TMEM79 VHLL CCT3 RHBG C1orf61 MEF2D IQGAP3 TTC24 APOA1BP GPATCH4 HAPLN2 BCAN NES CRABP2 ISG20L2 C1orf66 HDGF PRCC SH2D2A NTRK1 INSRR PEAR1 C1orf92 ARHGEF11 ETV3L

F


F


F

10

*

*

*

*

*

*


20

*

15

*

*

*

*

5

0

20

15

10

20 *

15

* * *

* *

5 q44

q42.11 q42.12 q42.13 q42.2 q42.3 q43

q41

q32.2 q32.3

q32.1

q31.1 q31.2 q31.3

q25.3

q21.1 q21.2 q21.3 q22 q23.1 q23.2 q23.3 q24.1 q24.2 q24.3 q25.1 q25.2

q12

p13.3 p13.2 p13.1 p12 p11.2 p11.1 q11

p22.3 p22.2 p22.1 p21.3 p21.2 p21.1

p31.1

p31.2

p32.3 p32.2 p32.1 p31.3

p34.2 p34.1 p33

p36.33 p36.32 p36.31 p36.23 p36.22 p36.21 p36.13 p36.12 p36.11 p35.3 p35.2 p35.1 p34.3


●

PAM50 region2 chr1:151857900−155336224 Subtypes Segregation − from S100A1 to ETV3L PAM50 Normal LumB LumA Her2 Basal

5

0


IntClust region2 chr1:151857900−155336224 Subtypes Segregation − from S100A1 to ETV3L IntClust 1 2 3 4 5 6 7 8 9 10

0


k−means region2 chr1:151857900−155336224 Subtypes Segregation − from S100A1 to ETV3L

k−means

9 8 7 6 5 4 3 2 1


MYBPH

ADORA1

MYOG

25

PPFIA4

MYBPH

ADORA1

MYOG

PPFIA4

25

TMEM183A

CYB5R1

ADIPOR1

KLHL12

RABIF

RP11−480I12.9

RP11−480I12.6

JARID1B

F

MYBPH

ADORA1

MYOG

PPFIA4

TMEM183A

CYB5R1

ADIPOR1

KLHL12

RABIF

RP11−480I12.9

RP11−480I12.6

JARID1B

F 25

TMEM183A

CYB5R1

ADIPOR1

KLHL12

RABIF

RP11−480I12.9

RP11−480I12.6

JARID1B

F

*

20

15

*

10

5

0 *

20

15

*

20

15 *

10

5 q44

q42.11 q42.12 q42.13 q42.2 q42.3 q43

q41

q32.2 q32.3

q32.1

q31.1 q31.2 q31.3

q25.3

q21.1 q21.2 q21.3 q22 q23.1 q23.2 q23.3 q24.1 q24.2 q24.3 q25.1 q25.2

q12

p13.3 p13.2 p13.1 p12 p11.2 p11.1 q11

p22.3 p22.2 p22.1 p21.3 p21.2 p21.1

p31.1

p31.2

p32.3 p32.2 p32.1 p31.3

p34.2 p34.1 p33

p36.33 p36.32 p36.31 p36.23 p36.22 p36.21 p36.13 p36.12 p36.11 p35.3 p35.2 p35.1 p34.3


●

PAM50 region3 chr1:200963155−201411565 Subtypes Segregation − from JARID1B to MYBPH PAM50 Normal LumB LumA Her2 Basal

10

5

0


IntClust region3 chr1:200963155−201411565 Subtypes Segregation − from JARID1B to MYBPH IntClust 1 2 3 4 5 6 7 8 9 10

0


k−means region3 chr1:200963155−201411565 Subtypes Segregation − from JARID1B to MYBPH

k−means

9 8 7 6 5 4 3 2 1


RAB7L1

NUCKS1

SLC45A3

ELK4

MFSD4

PCTK3

LEMD1

KLHDC8A

NUAK2

TMCC2

DSTYK

TMEM81

RBBP5

CNTN2

NFASC

LRRN2

MDM4

F

RAB7L1

NUCKS1

SLC45A3

ELK4

MFSD4

PCTK3

LEMD1

KLHDC8A

NUAK2

TMCC2

DSTYK

TMEM81

RBBP5

CNTN2

NFASC

LRRN2

MDM4

F

RAB7L1

NUCKS1

SLC45A3

ELK4

MFSD4

PCTK3

LEMD1

KLHDC8A

NUAK2

TMCC2

DSTYK

TMEM81

RBBP5

CNTN2

NFASC

LRRN2

MDM4

F

20

15

10

5

0

20

15

20

15

10

5 q44

q42.11 q42.12 q42.13 q42.2 q42.3 q43

q41

q32.2 q32.3

q32.1

q31.1 q31.2 q31.3

q25.3

q21.1 q21.2 q21.3 q22 q23.1 q23.2 q23.3 q24.1 q24.2 q24.3 q25.1 q25.2

q12

p13.3 p13.2 p13.1 p12 p11.2 p11.1 q11

p22.3 p22.2 p22.1 p21.3 p21.2 p21.1

p31.1

p31.2

p32.3 p32.2 p32.1 p31.3

p34.2 p34.1 p33

p36.33 p36.32 p36.31 p36.23 p36.22 p36.21 p36.13 p36.12 p36.11 p35.3 p35.2 p35.1 p34.3


●

PAM50 region4 chr1:202752134−204048784 Subtypes Segregation − from MDM4 to RAB7L1 PAM50 Normal LumB LumA Her2 Basal

10

5

0


IntClust region4 chr1:202752134−204048784 Subtypes Segregation − from MDM4 to RAB7L1 IntClust 1 2 3 4 5 6 7 8 9 10

0


k−means region4 chr1:202752134−204048784 Subtypes Segregation − from MDM4 to RAB7L1

k−means

9 8 7 6 5 4 3 2 1

q29

q26.32 q26.33 q27.1 q27.2 q27.3 q28

q26.31

q26.2

q26.1

q25.1 q25.2 q25.31 q25.32 q25.33

q24

q22.1 q22.2 q22.3 q23

q12.1 q12.2 q12.3 q13.11 q13.12 q13.13 q13.2 q13.31 q13.32 q13.33 q21.1 q21.2 q21.3

p12.2 p12.1 p11.2 p11.1 q11.1 q11.2

p12.3

p13

p14.1

p14.2

p21.2 p21.1 p14.3

p23 p22.3 p22.2 p22.1 p21.33 p21.32 p21.31

p24.1

p24.2

p24.3

p25.3 p25.2 p25.1

p26.3 p26.2 p26.1


●

PAM50 region5 chr3:150720722−151660305 Subtypes Segregation − from COMMD2 to TSC22D2 PAM50

8

Normal LumB LumA Her2 Basal

F

6 4 2

TSC22D2

PFN2

RNF13

COMMD2

0


IntClust region5 chr3:150720722−151660305 Subtypes Segregation − from COMMD2 to TSC22D2 IntClust

8

1 2 3 4 5 6 7 8 9 10

F

6 4 2

TSC22D2

PFN2

RNF13

COMMD2

0


k−means region5 chr3:150720722−151660305 Subtypes Segregation − from COMMD2 to TSC22D2 k−means

8

9 8 7 6 5 4 3 2 1

F

6 4 2


TSC22D2

PFN2

RNF13

COMMD2

0

AC092965.4−1 ZBBX SERPINI2 WDR49 PDCD10 SERPINI1 GOLIM4 C3orf50 EVI1 AC078802.14 MYNN LRRC34 LRRIQ4 LRRC31 SAMD7 SEC62 GPR160 PHC3 PRKCI SKIL CLDN11 SLC7A14 EIF5A2 SLC2A2 TNIK PLD1 GHSR TNFSF10 AADACL1 ECT2 SPATA16 NLGN1 NAALADL2 UBE2V1P2 KCNMB2 ZMAT3 PIK3CA KCNMB3 MFN1 GNB4 ACTL6A MRPL47 NDUFB5 USP13 PEX5L TTC14 FXR1 DNAJC19 SOX2 ATP11B DCUN1D1 MCCC1 LAMP3 MCF2L2 B3GNT5 KLHL6 KLHL24 YEATS2 MAP6D1 PARL ABCC5 HTR3D HTR3C HTR3E EIF2B5 DVL3 AP2M1 ABCF3 VWA5B2 ALG3 ECE2 CAMK2N2 PSMD2 EIF4G1 FAM131A CLCN2 POLR2H THPO CHRD EPHB3 MAGEF1 VPS8 C3orf70 EHHADH MAP3K13 TMEM41A LIPH SENP2 IGF2BP2 C3orf65 TRA2B ETV5 DGKG CRYGS TBCCD1 DNAJB11 AHSG FETUB HRG KNG1 EIF4A2 RFC4 ADIPOQ ST6GAL1 RPL39L RTP1 MASP1 RTP4 SST RTP2 BCL6 LPP TPRG1 TP63 LEPREL1 CLDN1 CLDN16 TMEM207 IL1RAP AC092952.4−2 OSTN UTS2D CCDC50 PYDC2

F


F


F


8

8

8

6

4

2

0 q29

q26.32 q26.33 q27.1 q27.2 q27.3 q28

q26.31

q26.2

q26.1

q25.1 q25.2 q25.31 q25.32 q25.33

q24

q22.1 q22.2 q22.3 q23

q12.1 q12.2 q12.3 q13.11 q13.12 q13.13 q13.2 q13.31 q13.32 q13.33 q21.1 q21.2 q21.3

p12.2 p12.1 p11.2 p11.1 q11.1 q11.2

p12.3

p13

p14.1

p14.2

p21.2 p21.1 p14.3

p23 p22.3 p22.2 p22.1 p21.33 p21.32 p21.31

p24.1

p24.2

p24.3

p25.3 p25.2 p25.1

p26.3 p26.2 p26.1


●

PAM50 region6 chr3:167464048−193320147 Subtypes Segregation − from AC092965.4−1 to PYDC2 PAM50

6 Normal LumB LumA Her2 Basal

4

2

0


IntClust region6 chr3:167464048−193320147 Subtypes Segregation − from AC092965.4−1 to PYDC2 IntClust

6

4

2 1 2 3 4 5 6 7 8 9 10

0


k−means region6 chr3:167464048−193320147 Subtypes Segregation − from AC092965.4−1 to PYDC2

k−means

9 8 7 6 5 4 3 2 1

AMTN


SDAD1

NAAA

PPEF2

USO1

G3BP2

CDKL2

C4orf26

THAP6

RCHY1

AC110760.3−2

BTC

AREGB

EREG

EPGN

MTHFD2L

CXCL2

CXCL3

CXCL5

PF4V1

CXCL6

IL8

RASSF6

AFM

AFP

ALB

ANKRD17

COX18

ADAMTS3

NPFFR2

GC

SLC4A4

DCK

MOBKL1A

GRSF1

RIPX

UTP3

IGJ

ENAM

AMBN

F

AMTN

SDAD1

NAAA

PPEF2

USO1

G3BP2

CDKL2

C4orf26

THAP6

RCHY1

AC110760.3−2

BTC

AREGB

EREG

EPGN

MTHFD2L

CXCL2

CXCL3

CXCL5

PF4V1

CXCL6

IL8

RASSF6

AFM

AFP

ALB

ANKRD17

COX18

ADAMTS3

NPFFR2

GC

SLC4A4

DCK

MOBKL1A

GRSF1

RIPX

UTP3

IGJ

ENAM

AMBN

F

AMTN

SDAD1

NAAA

PPEF2

USO1

G3BP2

CDKL2

C4orf26

THAP6

RCHY1

AC110760.3−2

BTC

AREGB

EREG

EPGN

MTHFD2L

CXCL2

CXCL3

CXCL5

PF4V1

CXCL6

IL8

RASSF6

AFM

AFP

ALB

ANKRD17

COX18

ADAMTS3

NPFFR2

GC

SLC4A4

DCK

MOBKL1A

GRSF1

RIPX

UTP3

IGJ

ENAM

AMBN

F

8

6

4

2

0

8

6

8

6

4

2 q35.2

q35.1

q33 q34.1 q34.2 q34.3

q32.3

q32.2

q32.1

q31.21 q31.22 q31.23 q31.3

q31.1

q28.3

q28.1 q28.2

q27

q26

q25

q24

q23

q21.1 q21.21 q21.22 q21.23 q21.3 q22.1 q22.2 q22.3

q13.3

q13.2

q13.1

q12

p12 p11 q11

p13

p14

p15.1

p15.2

p15.31

p15.32

p15.33

p16.3 p16.2 p16.1


●

PAM50 region7 chr4:71418887−77131137 Subtypes Segregation − from AMTN to SDAD1 PAM50 Normal LumB LumA Her2 Basal

4

2

0


IntClust region7 chr4:71418887−77131137 Subtypes Segregation − from AMTN to SDAD1 IntClust 1 2 3 4 5 6 7 8 9 10

0


k−means region7 chr4:71418887−77131137 Subtypes Segregation − from AMTN to SDAD1

k−means

9 8 7 6 5 4 3 2 1

q35.2

q35.1

q33 q34.1 q34.2 q34.3

q32.3

q32.2

q32.1

q31.21 q31.22 q31.23 q31.3

q31.1

q28.3

q28.1 q28.2

q27

q26

q25

q24

q23

q21.1 q21.21 q21.22 q21.23 q21.3 q22.1 q22.2 q22.3

q13.3

q13.2

q13.1

q12

p12 p11 q11

p13

p14

p15.1

p15.2

p15.31

p15.32

p15.33

p16.3 p16.2 p16.1

Chromosome 4 deletion outliers

●

PAM50 region8 chr4:184048238−184817325 Subtypes Segregation − from DCTD to RWDD4A 6

PAM50 Normal LumB LumA Her2 Basal

5

F

4 3 2 1

RWDD4A

ING2

CDKN2AIP

WWC2

DCTD

0


IntClust region8 chr4:184048238−184817325 Subtypes Segregation − from DCTD to RWDD4A 6

IntClust 1 2 3 4 5 6 7 8 9 10

5

F

4 3 2 1

RWDD4A

ING2

CDKN2AIP

WWC2

DCTD

0


k−means region8 chr4:184048238−184817325 Subtypes Segregation − from DCTD to RWDD4A 6

k−means 9 8 7 6 5 4 3 2 1

5

3 2 1


RWDD4A

ING2

CDKN2AIP

WWC2

0 DCTD

F

4


FOXO3

LACE1

SNX3

NR2E1

*

OSTM1

SEC63

SCML4

SOBP

PDSS2

BEND3

FOXO3

LACE1

SNX3

NR2E1

OSTM1

SEC63

SCML4

SOBP

PDSS2

BEND3

QRSL1

FOXO3

LACE1

SNX3

NR2E1

OSTM1

SEC63

SCML4

SOBP

PDSS2

BEND3

QRSL1

RTN4IP1

10

QRSL1

20 RTN4IP1

AIM1

ATG5

PRDM1

PREP

POPDC3

F 20

RTN4IP1

15 AIM1

ATG5

PRDM1

PREP

POPDC3

F 15

AIM1

10

ATG5

PRDM1

PREP

POPDC3

F

*

*

*

*

*

*

5

0 *

*

*

20

15

10

5 q27

q26

q25.3

q25.2

q25.1

q24.2 q24.3

q24.1

q23.3

q22.32 q22.33 q23.1 q23.2

q22.31

q22.1 q22.2

q21

q16.2 q16.3

q16.1

q15

q14.2 q14.3

q14.1

q13

q12

p12.2 p12.1 p11.2 p11.1 q11.1 q11.2

p12.3

p21.1

p21.2

p22.2 p22.1 p21.33 p21.32 p21.31

p22.3

p24.3 p24.2 p24.1 p23

p25.3 p25.2 p25.1


●

PAM50 region9 chr6:105712470−109112664 Subtypes Segregation − from POPDC3 to FOXO3 PAM50 Normal LumB LumA Her2 Basal

*

* * *

5

0


IntClust region9 chr6:105712470−109112664 Subtypes Segregation − from POPDC3 to FOXO3 IntClust 1 2 3 4 5 6 7 8 9 10

0


k−means region9 chr6:105712470−109112664 Subtypes Segregation − from POPDC3 to FOXO3

k−means

9 8 7 6 5 4 3 2 1

q36.3

q36.2

q36.1

q35

q34

q33

q31.32 q31.33 q32.1 q32.2 q32.3

q31.31

q31.2

q31.1

q22.2 q22.3

q22.1

q21.3

q21.12 q21.13 q21.2

q21.11

q11.23

q11.22

q11.21

p11.2 p11.1 q11.1

p12.3 p12.2 p12.1

p13

p14.1

p14.2

p14.3

p15.2 p15.1

p15.3

p21.1

p21.2

p21.3

p22.3 p22.2 p22.1


●

PAM50 region10 chr7:55054219−55545075 Subtypes Segregation − from EGFR to LANCL2 6


5

F

4 3 2 1

LANCL2

EGFR

0


IntClust region10 chr7:55054219−55545075 Subtypes Segregation − from EGFR to LANCL2 6

IntClust 1 2 3 4 5 6 7 8 9 10

5

F

4 3 2 1

LANCL2

EGFR

0


k−means region10 chr7:55054219−55545075 Subtypes Segregation − from EGFR to LANCL2 6

k−means 9 8 7 6 5 4 3 2 1

5

3 2 1

LANCL2

0 EGFR

F

4



PPP3CC

SLC39A14

PIWIL2

POLR3D

BMP1

SFTPC

LGI3

REEP4

HR

NUDT18

FAM160B2

EPB49

FGF17

NPM2

DOK2

GFRA2

LZTS1

ATP6V1B2

SLC18A1

LPL

20

INTS10

F

PPP3CC

SLC39A14

PIWIL2

POLR3D

BMP1

SFTPC

LGI3

REEP4

HR

NUDT18

FAM160B2

EPB49

FGF17

NPM2

DOK2

GFRA2

LZTS1

ATP6V1B2

SLC18A1

LPL

INTS10

F

PPP3CC

SLC39A14

PIWIL2

POLR3D

BMP1

SFTPC

LGI3

REEP4

HR

NUDT18

FAM160B2

EPB49

FGF17

NPM2

DOK2

GFRA2

LZTS1

ATP6V1B2

SLC18A1

LPL

INTS10

F

*

15

10

5

0

20

15

20

15

10

5 q24.3

q24.23

q24.22

q24.21

q24.13

q24.11 q24.12

q23.3

q23.2

q23.1

q22.3

q22.2

q22.1

q21.3

q21.2

q21.13

q21.12

q21.11

q13.3

q13.1 q13.2

q12.2 q12.3

q12.1

q11.21 q11.22 q11.23

p11.23 p11.22 p11.21 p11.1 q11.1

p12

p21.2 p21.1

p21.3

p22

p23.1

p23.2

p23.3


●

PAM50 region11 chr8:19719198−22454582 Subtypes Segregation − from INTS10 to PPP3CC PAM50 Normal LumB LumA Her2 Basal

10

5

0


IntClust region11 chr8:19719198−22454582 Subtypes Segregation − from INTS10 to PPP3CC IntClust 1 2 3 4 5 6 7 8 9 10

0


k−means region11 chr8:19719198−22454582 Subtypes Segregation − from INTS10 to PPP3CC

k−means

9 8 7 6 5 4 3 2 1

Genes & Chi−square Test INTS9

EXTL3

FZD3

FBXO16

ZNF395

PNOC

ELP3

C8orf80

SCARA5

PBK

ESCO2

CCDC25

SCARA3

CLU

EPHX2

CHRNA2

PTK2B

TRIM35

STMN4

ADRA1A

DPYSL2

PNMA2

BNIP3L

PPP2R2A

F 15

INTS9

EXTL3

FZD3

FBXO16

ZNF395

PNOC

ELP3

C8orf80

SCARA5

PBK

ESCO2

CCDC25

SCARA3

CLU

EPHX2

CHRNA2

PTK2B

TRIM35

STMN4

ADRA1A

DPYSL2

PNMA2

BNIP3L

PPP2R2A

F

INTS9

EXTL3

FZD3

FBXO16

ZNF395

PNOC

ELP3

C8orf80

SCARA5

PBK

ESCO2

CCDC25

SCARA3

CLU

EPHX2

CHRNA2

PTK2B

TRIM35

STMN4

ADRA1A

DPYSL2

PNMA2

BNIP3L

PPP2R2A

F

20

*

15

10

5

0

20

15

20 *

*

10

5 q24.3

q24.23

q24.22

q24.21

q24.13

q24.11 q24.12

q23.3

q23.2

q23.1

q22.3

q22.2

q22.1

q21.3

q21.2

q21.13

q21.12

q21.11

q13.3

q13.1 q13.2

q12.2 q12.3

q12.1

q11.21 q11.22 q11.23

p11.23 p11.22 p11.21 p11.1 q11.1

p12

p21.2 p21.1

p21.3

p22

p23.1

p23.2

p23.3


●

PAM50 region12 chr8:26204951−28803398 Subtypes Segregation − from PPP2R2A to INTS9 PAM50 Normal LumB LumA Her2 Basal

10

5

0


IntClust region12 chr8:26204951−28803398 Subtypes Segregation − from PPP2R2A to INTS9 IntClust 1 2 3 4 5 6 7 8 9 10

0


k−means region12 chr8:26204951−28803398 Subtypes Segregation − from PPP2R2A to INTS9

k−means

9 8 7 6 5 4 3 2 1

Genes & Chi−square Test ADAM32

*

ADAM32

ADAM9

TM2D2

HTRA4

*

ADAM9

TM2D2

HTRA4

PLEKHA2

20

PLEKHA2

TACC1

*

TACC1

AC069120.9

C8orf86

FGFR1

LETM2

*

AC069120.9

30

C8orf86

* WHSC1L1

30

FGFR1

* PPAPDC1B

*

LETM2

BAG4

*

WHSC1L1

* DDHD2

*

PPAPDC1B

*

BAG4

*

DDHD2

* LSM1

STAR

*

LSM1

* ASH2L

10

STAR

40 EIF4EBP1

*

ASH2L

ADRB3

GOT1L1

RAB11FIP1

40

EIF4EBP1

ADRB3

GOT1L1

RAB11FIP1

* BRF2

GPR124

PROSC *

BRF2

60

GPR124

* ERLIN2

60

PROSC

50 *

ERLIN2

ZNF703

F 50

ZNF703

F

ADAM32

ADAM9

TM2D2

HTRA4

PLEKHA2

TACC1

AC069120.9

C8orf86

FGFR1

LETM2

WHSC1L1

PPAPDC1B

DDHD2

BAG4

LSM1

STAR

ASH2L

EIF4EBP1

ADRB3

GOT1L1

RAB11FIP1

BRF2

GPR124

PROSC

ERLIN2

ZNF703

F

70

*

*

*

*

*

*

20

10

0

70

60

50

40

70

* *

* * *

* * *

*

q24.3

q24.23

q24.22

q24.21

q24.13

q24.11 q24.12

q23.3

q23.2

q23.1

q22.3

q22.2

q22.1

q21.3

q21.2

q21.13

q21.12

q21.11

q13.3

q13.1 q13.2

q12.2 q12.3

q12.1

q11.21 q11.22 q11.23

p11.23 p11.22 p11.21 p11.1 q11.1

p12

p21.2 p21.1

p21.3

p22

p23.1

p23.2

p23.3


●

PAM50 region13 chr8:37672459−39261593 Subtypes Segregation − from ZNF703 to ADAM32 PAM50 Normal LumB LumA Her2 Basal

30

20

10

0


IntClust region13 chr8:37672459−39261593 Subtypes Segregation − from ZNF703 to ADAM32 IntClust 1 2 3 4 5 6 7 8 9 10

0 *


k−means region13 chr8:37672459−39261593 Subtypes Segregation − from ZNF703 to ADAM32

k−means

9 8 7 6 5 4 3 2 1


ZFAT

* ZFAT

ST3GAL1

*

ST3GAL1

NDRG1

WISP1

SLA

*

NDRG1

WISP1

SLA

TG

10

TG

30 PHF20L1

TMEM71

LRRC6

KCNQ3

OC90

30

PHF20L1

TMEM71

LRRC6

10

KCNQ3

*

OC90

*

EFR3A

ADCY8

ASAP1

FAM49B

*

EFR3A

ADCY8

ASAP1

FAM49B

20 GSDMC

TMEM75

20

GSDMC

40

TMEM75

40

AC084123.9

MYC

POU5F1P1

FAM84B

F

ZFAT

ST3GAL1

NDRG1

WISP1

SLA

TG

PHF20L1

TMEM71

LRRC6

KCNQ3

OC90

EFR3A

ADCY8

ASAP1

FAM49B

GSDMC

TMEM75

AC084123.9

MYC

POU5F1P1

FAM84B

F 40

AC084123.9

MYC

POU5F1P1

FAM84B

F

*

*

*

*

*

*

0 *

30 *

*

*

* *

*

*

q24.3

q24.23

q24.22

q24.21

q24.13

q24.11 q24.12

q23.3

q23.2

q23.1

q22.3

q22.2

q22.1

q21.3

q21.2

q21.13

q21.12

q21.11

q13.3

q13.1 q13.2

q12.2 q12.3

q12.1

q11.21 q11.22 q11.23

p11.23 p11.22 p11.21 p11.1 q11.1

p12

p21.2 p21.1

p21.3

p22

p23.1

p23.2

p23.3


●

PAM50 region14 chr8:127300001−136500000 Subtypes Segregation − from FAM84B to ZFAT PAM50 Normal LumB LumA Her2 Basal

20

10

0


IntClust region14 chr8:127300001−136500000 Subtypes Segregation − from FAM84B to ZFAT IntClust 1 2 3 4 5 6 7 8 9 10

0


k−means region14 chr8:127300001−136500000 Subtypes Segregation − from FAM84B to ZFAT

k−means

9 8 7 6 5 4 3 2 1

q34.3

q34.11 q34.12 q34.13 q34.2

q33.3

q33.2

q33.1

q32

q31.3

q31.2

q31.1

q22.33

q22.32

q21.33 q22.1 q22.2 q22.31

q21.32

q21.2 q21.31

q21.13

q13 q21.11 q21.12

q12

q11

p11.1

p12 p11.2

p13.2 p13.1

p13.3

p21.1

p21.2

p21.3

p22.3 p22.2 p22.1

p23

p24.1

p24.3 p24.2


●

PAM50 region15 chr9:21507627−22140976 Subtypes Segregation − from MTAP to CDKN2B PAM50

15


F

10

5

CDKN2B

CDKN2A

MTAP

0


IntClust region15 chr9:21507627−22140976 Subtypes Segregation − from MTAP to CDKN2B IntClust

15

1 2 3 4 5 6 7 8 9 10

F

10

5

CDKN2B

CDKN2A

MTAP

0


k−means region15 chr9:21507627−22140976 Subtypes Segregation − from MTAP to CDKN2B k−means

15

9 8 7 6 5 4 3 2 1

F

10

5

CDKN2B

CDKN2A

MTAP

0


q26.3

q26.2

q26.13

q26.11 q26.12

q25.3

q25.2

q25.1

q24.1 q24.2 q24.31 q24.32 q24.33

q23.31 q23.32 q23.33

q23.2

q23.1

q22.3

q22.2

q22.1

q21.3

q21.2

q21.1

q11.23

q11.22

q11.21

p11.1 q11.1

p11.21

p11.22

p11.23

p12.1

p12.2

p12.33 p12.32 p12.31

p13

p14

p15.3 p15.2 p15.1


●

PAM50 region16 chr10:30702999−32803238 Subtypes Segregation − from RP11−730A19.6 to FAM107B PAM50

14 12


*

F

10 *

8 6 4 2

FAM107B

FRMD4A

PRPF18

BEND7

SEPHS1

PHYH

C10orf49

MCM10

RP11−730A19.6

0


IntClust region16 chr10:30702999−32803238 Subtypes Segregation − from RP11−730A19.6 to FAM107B IntClust

14 12

* *

10 F

1 2 3 4 5 6 7 8 9 10

*

*

8 6 4 2

FAM107B

FRMD4A

PRPF18

BEND7

SEPHS1

PHYH

C10orf49

MCM10

RP11−730A19.6

0


k−means region16 chr10:30702999−32803238 Subtypes Segregation − from RP11−730A19.6 to FAM107B k−means

14 12

9 8 7 6 5 4 3 2 1

*

8 6 4 2


FAM107B

FRMD4A

PRPF18

BEND7

SEPHS1

PHYH

C10orf49

MCM10

0 RP11−730A19.6

F

10


RP11−369J21.6

RP11−369J21.4

SFTPD

AL512662.8−3

AL512662.8−2

AL135925.10

FAM22C

SFTPA1B

FAM22B

SFTPA1

ZCCHC24

PPIF

ZMIZ1

RPS24

POLR3A

DLG5

KCNMA1

C10orf11

ZNF503

COMTD1

F

RP11−369J21.6

RP11−369J21.4

SFTPD

AL512662.8−3

AL512662.8−2

AL135925.10

FAM22C

SFTPA1B

FAM22B

SFTPA1

ZCCHC24

PPIF

ZMIZ1

RPS24

POLR3A

DLG5

KCNMA1

C10orf11

ZNF503

COMTD1

F

RP11−369J21.6

RP11−369J21.4

SFTPD

AL512662.8−3

AL512662.8−2

AL135925.10

FAM22C

SFTPA1B

FAM22B

SFTPA1

ZCCHC24

PPIF

ZMIZ1

RPS24

POLR3A

DLG5

KCNMA1

C10orf11

ZNF503

COMTD1

F 20

15

20

20

15

10

5

0 q26.3

q26.2

q26.13

q26.11 q26.12

q25.3

q25.2

q25.1

q24.1 q24.2 q24.31 q24.32 q24.33

q23.31 q23.32 q23.33

q23.2

q23.1

q22.3

q22.2

q22.1

q21.3

q21.2

q21.1

q11.23

q11.22

q11.21

p11.1 q11.1

p11.21

p11.22

p11.23

p12.1

p12.2

p12.33 p12.32 p12.31

p13

p14

p15.3 p15.2 p15.1


●

PAM50 region17 chr10:76663735−81842287 Subtypes Segregation − from COMTD1 to RP11−369J21.6 PAM50 Normal LumB LumA Her2 Basal

10

5

0


IntClust region17 chr10:76663735−81842287 Subtypes Segregation − from COMTD1 to RP11−369J21.6 IntClust

15

10

5 1 2 3 4 5 6 7 8 9 10

0


k−means region17 chr10:76663735−81842287 Subtypes Segregation − from COMTD1 to RP11−369J21.6

k−means

9 8 7 6 5 4 3 2 1

WAPAL

LIPJ


IFIT5

IFIT1

IFIT1L

IFIT3

IFIT2

LIPA

CH25H

FAS

ACTA2

STAMBPL1

ANKRD22

LIPM

LIPN

LIPK

LIPF

30 IFIT5

IFIT1

IFIT1L

IFIT3

IFIT2

LIPA

CH25H

FAS

ACTA2

STAMBPL1

ANKRD22

LIPM

LIPN

LIPK

LIPF

LIPJ

C10orf59

PTEN

AFDC1

PAPSS2

MINPP1

FAM22D

FAM35A

GLUD1

FAM25A

AGAP11

AL136982.11−2

C10orf116

SNCG

MMRN2

BMPR1A

LDB3

OPN4

30

C10orf59

WAPAL RPL7AP8

F

WAPAL

IFIT5

IFIT1

IFIT1L

IFIT3

IFIT2

LIPA

CH25H

FAS

ACTA2

STAMBPL1

ANKRD22

LIPM

LIPN

LIPK

LIPF

LIPJ

C10orf59

PTEN

AFDC1

PAPSS2

MINPP1

FAM22D

FAM35A

GLUD1

FAM25A

AGAP11

AL136982.11−2

C10orf116

SNCG

MMRN2

BMPR1A

LDB3

OPN4

RPL7AP8

F 30

PTEN

AFDC1

PAPSS2

MINPP1

FAM22D

FAM35A

GLUD1

FAM25A

AGAP11

AL136982.11−2

C10orf116

SNCG

MMRN2

BMPR1A

LDB3

OPN4

RPL7AP8

F

*

25

20

15

10

5

0 *

25

20

*

25

20

15

10

5 q26.3

q26.2

q26.13

q26.11 q26.12

q25.3

q25.2

q25.1

q24.1 q24.2 q24.31 q24.32 q24.33

q23.31 q23.32 q23.33

q23.2

q23.1

q22.3

q22.2

q22.1

q21.3

q21.2

q21.1

q11.23

q11.22

q11.21

p11.1 q11.1

p11.21

p11.22

p11.23

p12.1

p12.2

p12.33 p12.32 p12.31

p13

p14

p15.3 p15.2 p15.1


●

PAM50 region18 chr10:88085774−91245913 Subtypes Segregation − from WAPAL to IFIT5 PAM50 Normal LumB LumA Her2 Basal

15

10

5

0


IntClust region18 chr10:88085774−91245913 Subtypes Segregation − from WAPAL to IFIT5 IntClust 1 2 3 4 5 6 7 8 9 10

0


k−means region18 chr10:88085774−91245913 Subtypes Segregation − from WAPAL to IFIT5

k−means

9 8 7 6 5 4 3 2 1


C11orf55

LDLRAD3

TRIM44

AC090625.5−2

SLC1A2

CD44

PDHX

APIP

EHF

6 C11orf55

LDLRAD3

TRIM44

AC090625.5−2

SLC1A2

CD44

PDHX

APIP

EHF

ELF5

CAT

ABTB2

NAT10

CAPRIN1

FBXO3

CD59

C11orf91

C11orf41

HIPK3

CSTF3

TCP11L1

DEPDC7

QSER1

PRRG4

CCDC73

EIF3M

WIT1

WT1

RCN1

PAX6

ELP4

IMMP1L

DNAJC24

F

C11orf55

LDLRAD3

TRIM44

AC090625.5−2

SLC1A2

CD44

PDHX

APIP

EHF

ELF5

CAT

ABTB2

NAT10

CAPRIN1

FBXO3

CD59

C11orf91

C11orf41

HIPK3

CSTF3

TCP11L1

DEPDC7

QSER1

PRRG4

CCDC73

EIF3M

WIT1

WT1

RCN1

PAX6

ELP4

IMMP1L

DNAJC24

6

ELF5

CAT

ABTB2

NAT10

CAPRIN1

FBXO3

CD59

C11orf91

C11orf41

HIPK3

CSTF3

TCP11L1

DEPDC7

QSER1

PRRG4

CCDC73

EIF3M

WIT1

WT1

RCN1

PAX6

ELP4

IMMP1L

DNAJC24

F

F

14

12

10

8

4

*

2

0

14 q25

q24.3

q24.2

q24.1

q23.3

q23.1 q23.2

q22.3

q22.2

q22.1

q21

q14.3

q14.2

q14.1

q13.5

q13.4

q13.2 q13.3

q12.1 q12.2 q12.3 q13.1

q11

p11.12 p11.11

p11.2

p12

p13

p14.2 p14.1

p14.3

p15.1

p15.2

p15.3

p15.4

p15.5


●

PAM50 region19 chr11:31347953−36251485 Subtypes Segregation − from DNAJC24 to C11orf55

12 PAM50

10


4 *

2

0


IntClust region19 chr11:31347953−36251485 Subtypes Segregation − from DNAJC24 to C11orf55

14

12 IntClust

10

8

6

4

2 1 2 3 4 5 6 7 8 9 10

0


k−means region19 chr11:31347953−36251485 Subtypes Segregation − from DNAJC24 to C11orf55

k−means

9 8 7 6 5 4 3 2 1

q25

q24.3

q24.2

q24.1

q23.3

q23.1 q23.2

q22.3

q22.2

q22.1

q21

q14.3

q14.2

q14.1

q13.5

q13.4

q13.2 q13.3

q12.1 q12.2 q12.3 q13.1

q11

p11.12 p11.11

p11.2

p12

p13

p14.2 p14.1

p14.3

p15.1

p15.2

p15.3

p15.4

p15.5


●

PAM50 region20 chr11:65576392−66471916 Subtypes Segregation − from RBM4 to LRFN4 PAM50

30


*

25 F

20 *

15 10 5 LRFN4

PC

RCE1

C11orf80

SPTBN2

RBM4B

TMEM137

RBM4

0


IntClust region20 chr11:65576392−66471916 Subtypes Segregation − from RBM4 to LRFN4 IntClust

30

*

25 F

20 *

15

* *

10

* *

5

1 2 3 4 5 6 7 8 9 10

LRFN4

PC

RCE1

C11orf80

SPTBN2

RBM4B

TMEM137

RBM4

0


k−means region20 chr11:65576392−66471916 Subtypes Segregation − from RBM4 to LRFN4 k−means

30

9 8 7 6 5 4 3 2 1

*

25 F

20 15 10 5


LRFN4

PC

RCE1

C11orf80

SPTBN2

RBM4B

TMEM137

RBM4

0

Genes & Chi−square Test *

*

* CTTN

*

PPFIA1

CTTN

PPFIA1

*

CTTN

80 FADD

*

PPFIA1

80

FADD

ANO1

FGF3

FGF4

FGF19

ORAOV1

AP001888.4

CCND1

MYEOV

80

FADD

ANO1

FGF3

FGF4

FGF19

ORAOV1

AP001888.4

*

ANO1

FGF3

FGF4

FGF19

ORAOV1

*

AP001888.4

*

CCND1

MYEOV

TPCN2

MRGPRF

MRGPRD

IGHMBP2

MRPL21

CPT1A

MTL5

F 40

CCND1

MYEOV

TPCN2

MRGPRF

MRGPRD

IGHMBP2

*

TPCN2

MRGPRF

MRGPRD

IGHMBP2

40 MRPL21

F 40

MRPL21

CPT1A

MTL5

20

CPT1A

MTL5

F

*

*

60

*

20

0 * *

60

* *

60

* * *

* *

q25

q24.3

q24.2

q24.1

q23.3

q23.1 q23.2

q22.3

q22.2

q22.1

q21

q14.3

q14.2

q14.1

q13.5

q13.4

q13.2 q13.3

q12.1 q12.2 q12.3 q13.1

q11

p11.12 p11.11

p11.2

p12

p13

p14.2 p14.1

p14.3

p15.1

p15.2

p15.3

p15.4

p15.5


●

PAM50 region21 chr11:68231484−70185520 Subtypes Segregation − from MTL5 to CTTN PAM50 Normal LumB LumA Her2 Basal

*

20

0


IntClust region21 chr11:68231484−70185520 Subtypes Segregation − from MTL5 to CTTN IntClust 1 2 3 4 5 6 7 8 9 10

0 *


k−means region21 chr11:68231484−70185520 Subtypes Segregation − from MTL5 to CTTN

k−means

9 8 7 6 5 4 3 2 1

q25

q24.3

q24.2

q24.1

q23.3

q23.1 q23.2

q22.3

q22.2

q22.1

q21

q14.3

q14.2

q14.1

q13.5

q13.4

q13.2 q13.3

q12.1 q12.2 q12.3 q13.1

q11

p11.12 p11.11

p11.2

p12

p13

p14.2 p14.1

p14.3

p15.1

p15.2

p15.3

p15.4

p15.5


●

PAM50 region22 chr11:71578255−72747564 Subtypes Segregation − from FOLR2 to P2RY6 PAM50 Normal LumB LumA Her2 Basal

F

15

10

5

P2RY6

P2RY2

FCHSD2

ATG16L2

STARD10

ARAP1

PDE2A

CLPB

PHOX2A

INPPL1

FOLR2

0


IntClust region22 chr11:71578255−72747564 Subtypes Segregation − from FOLR2 to P2RY6 IntClust *

15

1 2 3 4 5 6 7 8 9 10

F

*

*

10 *

5

P2RY6

P2RY2

FCHSD2

ATG16L2

STARD10

ARAP1

PDE2A

CLPB

PHOX2A

INPPL1

FOLR2

0


k−means region22 chr11:71578255−72747564 Subtypes Segregation − from FOLR2 to P2RY6 k−means 9 8 7 6 5 4 3 2 1

10

5


P2RY6

P2RY2

FCHSD2

ATG16L2

STARD10

ARAP1

PDE2A

CLPB

PHOX2A

INPPL1

0 FOLR2

F

15

Genes & Chi−square Test GAB2

*

GAB2

*

USP35

KCTD21

*

USP35

KCTD21

*

ALG8

NDUFC2

GAB2

USP35

KCTD21

ALG8

NDUFC2

THRSP

KCTD14

INTS4

C11orf67

*

ALG8

50

NDUFC2

THRSP

*

THRSP

KCTD14

*

KCTD14

INTS4

*

INTS4

C11orf67

RSF1

CLNS1A

AQP11

PAK1

60

C11orf67

RSF1

50

RSF1

40 CLNS1A

60

CLNS1A

*

AQP11

PAK1

GDPD4

MYO7A

CAPN5

B3GNT6

ACER3

TSKU

LRRC32

C11orf30

PRKRIR

WNT11

UVRAG

F 40

AQP11

PAK1

GDPD4

10

GDPD4

MYO7A

CAPN5

B3GNT6

*

MYO7A

CAPN5

B3GNT6

ACER3

*

ACER3

TSKU

LRRC32

C11orf30

20

TSKU

LRRC32

C11orf30

PRKRIR

WNT11

F 40

PRKRIR

WNT11

UVRAG

30

UVRAG

F

70

60

*

*

30

20

10

0

70

*

50

*

70

*

* *

* *

*

* *

* *

q25

q24.3

q24.2

q24.1

q23.3

q23.1 q23.2

q22.3

q22.2

q22.1

q21

q14.3

q14.2

q14.1

q13.5

q13.4

q13.2 q13.3

q12.1 q12.2 q12.3 q13.1

q11

p11.12 p11.11

p11.2

p12

p13

p14.2 p14.1

p14.3

p15.1

p15.2

p15.3

p15.4

p15.5


●

PAM50 region23 chr:75203860−77963367 Subtypes Segregation − from UVRAG to GAB2 PAM50 Normal LumB LumA Her2 Basal

30

20

10

0


IntClust region23 chr:75203860−77963367 Subtypes Segregation − from UVRAG to GAB2 IntClust 1 2 3 4 5 6 7 8 9 10

0


k−means region23 chr:75203860−77963367 Subtypes Segregation − from UVRAG to GAB2

k−means

9 8 7 6 5 4 3 2 1


RAD51AP1

AC008012.8

FGF6

FGF23

C12orf5

CCND2

PARP11

EFCAB4B

PRMT8

TSPAN9

* RAD51AP1

AC008012.8

FGF6

FGF23

C12orf5

CCND2

PARP11

EFCAB4B

PRMT8

TSPAN9

TEAD4

TULP3

C12orf32

FOXM1

NRIP2

FKBP4

CACNA1C

DCP1B

CACNA2D4

ADIPOR2

WNT5B

FBXL14

RAD51AP1

AC008012.8

FGF6

FGF23

C12orf5

CCND2

PARP11

EFCAB4B

PRMT8

TSPAN9

TEAD4

TULP3

C12orf32

FOXM1

NRIP2

FKBP4

CACNA1C

DCP1B

CACNA2D4

ADIPOR2

WNT5B

FBXL14

ERC1

RAD52

WNK1

NINJ2

B4GALNT3

*

TEAD4

TULP3

C12orf32

FOXM1

NRIP2

FKBP4

CACNA1C

DCP1B

CACNA2D4

ADIPOR2

WNT5B

FBXL14

6 ERC1

RAD52

WNK1

NINJ2

B4GALNT3

CCDC77

6

ERC1

RAD52

WNK1

NINJ2

8

B4GALNT3

10 CCDC77

JARID1A

F 10

CCDC77

JARID1A

F 8

JARID1A

F

12

*

*

*

*

*

4

2

0

12

* *

* *

12

10

8

6

4

2 q24.33

q24.32

q24.31

q24.11 q24.12 q24.13 q24.21 q24.22 q24.23

q23.3

q23.2

q23.1

q22

q21.33

q21.32

q21.31

q21.2

q21.1

q15

q14.2 q14.3

q14.1

q13.2 q13.3

q13.13

q13.12

q13.11

q12

p11.21 p11.1 q11

p11.23 p11.22

p12.1

p12.2

p12.3

p13.2 p13.1

p13.31

p13.32

p13.33


●

PAM50 region24 chr12:259484−4539475 Subtypes Segregation − from JARID1A to RAD51AP1 PAM50 Normal LumB LumA Her2 Basal

4

2

0


IntClust region24 chr12:259484−4539475 Subtypes Segregation − from JARID1A to RAD51AP1 IntClust 1 2 3 4 5 6 7 8 9 10

0


k−means region24 chr12:259484−4539475 Subtypes Segregation − from JARID1A to RAD51AP1

k−means

9 8 7 6 5 4 3 2 1


CNOT2

C12orf28

*

RAB3IP

BEST3

CCT2

AC018921.22

15

FRS2

YEATS4

LYZ

CPM

CNOT2

C12orf28

RAB3IP

BEST3

CCT2

AC018921.22

FRS2

YEATS4

LYZ

CPM

MDM2

SLC35E3

15

MDM2

SLC35E3

20 NUP107

AC015550.19−1

MDM1

F 20

NUP107

AC015550.19−1

MDM1

F

CNOT2

C12orf28

RAB3IP

BEST3

CCT2

AC018921.22

FRS2

YEATS4

LYZ

CPM

MDM2

SLC35E3

NUP107

AC015550.19−1

MDM1

F

*

*

*

*

*

10

5

0

20

15

*

*

10

5 q24.33

q24.32

q24.31

q24.11 q24.12 q24.13 q24.21 q24.22 q24.23

q23.3

q23.2

q23.1

q22

q21.33

q21.32

q21.31

q21.2

q21.1

q15

q14.2 q14.3

q14.1

q13.2 q13.3

q13.13

q13.12

q13.11

q12

p11.21 p11.1 q11

p11.23 p11.22

p12.1

p12.2

p12.3

p13.2 p13.1

p13.31

p13.32

p13.33


●

PAM50 region25 chr12:66974613−69035040 Subtypes Segregation − from MDM1 to CNOT2 PAM50 Normal LumB LumA Her2 Basal

10

5

0


IntClust region25 chr12:66974613−69035040 Subtypes Segregation − from MDM1 to CNOT2 IntClust 1 2 3 4 5 6 7 8 9 10

0


k−means region25 chr12:66974613−69035040 Subtypes Segregation − from MDM1 to CNOT2

k−means

9 8 7 6 5 4 3 2 1

q34

q33.3

q33.2

q33.1

q32.2 q32.3

q32.1

q31.3

q31.2

q31.1

q22.1 q22.2 q22.3

q21.33

q21.32

q21.31

q21.2

q21.1

q14.3

q14.2

q14.12 q14.13

q14.11

q13.3

q13.1 q13.2

q12.3

q12.13 q12.2

q12.12

q12.11

q11

p11.1

p11.2

p12

p13


●

PAM50 region26 chr13:37224832−39767421 Subtypes Segregation − from UFM1 to COG6 PAM50

8


F

6 4 2

COG6

LHFP

NHLRC3

C13orf23

STOML3

FREM2

UFM1

0


IntClust region26 chr13:37224832−39767421 Subtypes Segregation − from UFM1 to COG6 IntClust

8

1 2 3 4 5 6 7 8 9 10

F

6 4 2

COG6

LHFP

NHLRC3

C13orf23

STOML3

FREM2

UFM1

0


k−means region26 chr13:37224832−39767421 Subtypes Segregation − from UFM1 to COG6 k−means

8

9 8 7 6 5 4 3 2 1

F

6 4 2


COG6

LHFP

NHLRC3

C13orf23

STOML3

FREM2

UFM1

0

SIAH3


RNASEH2B

RP11−480P3.1

DLEU1

KCNRG

TRIM13

C13orf1

KPNA3

EBPL

ARL11

AL135901.23−1

RCBTB1

PHF11

SETDB2

CAB39L

CDADC1

MLNR

FNDC3A

CYSLTR2

RCBTB2

P2RY5

RB1

ITM2B

MED4

NUDT15

SUCLA2

HTR2A

ESD

LRCH1

PPP1R2P4

LRRC63

LCP1

CPB2

ZC3H13

F

SIAH3

RNASEH2B

RP11−480P3.1

DLEU1

KCNRG

TRIM13

C13orf1

KPNA3

EBPL

ARL11

AL135901.23−1

RCBTB1

PHF11

SETDB2

CAB39L

CDADC1

MLNR

FNDC3A

CYSLTR2

RCBTB2

P2RY5

RB1

ITM2B

MED4

NUDT15

SUCLA2

HTR2A

ESD

LRCH1

PPP1R2P4

LRRC63

LCP1

CPB2

ZC3H13

F

SIAH3

RNASEH2B

RP11−480P3.1

DLEU1

KCNRG

TRIM13

C13orf1

KPNA3

EBPL

ARL11

AL135901.23−1

RCBTB1

PHF11

SETDB2

CAB39L

CDADC1

MLNR

FNDC3A

CYSLTR2

RCBTB2

P2RY5

RB1

ITM2B

MED4

NUDT15

SUCLA2

HTR2A

ESD

LRCH1

PPP1R2P4

LRRC63

LCP1

CPB2

ZC3H13

F

8

6

4

2

0 q34

q33.3

q33.2

q33.1

q32.2 q32.3

q32.1

q31.3

q31.2

q31.1

q22.1 q22.2 q22.3

q21.33

q21.32

q21.31

q21.2

q21.1

q14.3

q14.2

q14.12 q14.13

q14.11

q13.3

q13.1 q13.2

q12.3

q12.13 q12.2

q12.12

q12.11

q11

p11.1

p11.2

p12

p13


●

PAM50 region27 chr13:45192796−50537115 Subtypes Segregation − from SIAH3 to RNASEH2B

8 PAM50

6


2

0


IntClust region27 chr13:45192796−50537115 Subtypes Segregation − from SIAH3 to RNASEH2B

8 IntClust

6

4

2 1 2 3 4 5 6 7 8 9 10

0


k−means region27 chr13:45192796−50537115 Subtypes Segregation − from SIAH3 to RNASEH2B

k−means

9 8 7 6 5 4 3 2 1

q34

q33.3

q33.2

q33.1

q32.2 q32.3

q32.1

q31.3

q31.2

q31.1

q22.1 q22.2 q22.3

q21.33

q21.32

q21.31

q21.2

q21.1

q14.3

q14.2

q14.12 q14.13

q14.11

q13.3

q13.1 q13.2

q12.3

q12.13 q12.2

q12.12

q12.11

q11

p11.1

p11.2

p12

p13


●

PAM50 region28 chr13:49235120−49745329 Subtypes Segregation − from C13orf1 to DLEU1 8


F

6 4 2

DLEU1

KCNRG

TRIM13

C13orf1

0


IntClust region28 chr13:49235120−49745329 Subtypes Segregation − from C13orf1 to DLEU1 8

IntClust 1 2 3 4 5 6 7 8 9 10

F

6 4 2

DLEU1

KCNRG

TRIM13

C13orf1

0


k−means region28 chr13:49235120−49745329 Subtypes Segregation − from C13orf1 to DLEU1 8

k−means 9 8 7 6 5 4 3 2 1

4 2


DLEU1

KCNRG

TRIM13

0 C13orf1

F

6

q34

q33.3

q33.2

q33.1

q32.2 q32.3

q32.1

q31.3

q31.2

q31.1

q22.1 q22.2 q22.3

q21.33

q21.32

q21.31

q21.2

q21.1

q14.3

q14.2

q14.12 q14.13

q14.11

q13.3

q13.1 q13.2

q12.3

q12.13 q12.2

q12.12

q12.11

q11

p11.1

p11.2

p12

p13


●

PAM50 region29 chr13:71316783−73426040 Subtypes Segregation − from C13orf37 to KLF5 8


F

6 4 2

KLF5

PIBF1

DIS3

C13orf34

C13orf37

0


IntClust region29 chr13:71316783−73426040 Subtypes Segregation − from C13orf37 to KLF5 8

IntClust 1 2 3 4 5 6 7 8 9 10

F

6 4 2

KLF5

PIBF1

DIS3

C13orf34

C13orf37

0


k−means region29 chr13:71316783−73426040 Subtypes Segregation − from C13orf37 to KLF5 8

k−means 9 8 7 6 5 4 3 2 1

4 2


KLF5

PIBF1

DIS3

C13orf34

0 C13orf37

F

6


TTBK2

CDAN1

STARD9

AC090510.4−2

CEP27

LRRC57

SNAP23

ZFP106

CAPN3

GANC

TMEM87A

VPS39

PLA2G4F

PLA2G4D

PLA2G4E

F

TTBK2

CDAN1

STARD9

AC090510.4−2

CEP27

LRRC57

SNAP23

ZFP106

CAPN3

GANC

TMEM87A

VPS39

PLA2G4F

PLA2G4D

PLA2G4E

F

TTBK2

CDAN1

STARD9

AC090510.4−2

CEP27

LRRC57

SNAP23

ZFP106

CAPN3

GANC

TMEM87A

VPS39

PLA2G4F

PLA2G4D

PLA2G4E

F

8

6

4

2

0

8

6

8

6

4

2 q26.3

q26.2

q26.1

q25.3

q25.2

q25.1

q24.2 q24.3

q24.1

q23

q22.31 q22.32 q22.33

q22.2

q22.1

q21.3

q21.2

q21.1

q15.1 q15.2 q15.3

q14

q13.1 q13.2 q13.3

q12

q11.2

q11.1

p11.1

p11.2

p12

p13


●

PAM50 region30 chr15:40032508−41155167 Subtypes Segregation − from PLA2G4E to TTBK2 PAM50 Normal LumB LumA Her2 Basal

4

2

0


IntClust region30 chr15:40032508−41155167 Subtypes Segregation − from PLA2G4E to TTBK2 IntClust 1 2 3 4 5 6 7 8 9 10

0


k−means region30 chr15:40032508−41155167 Subtypes Segregation − from PLA2G4E to TTBK2

k−means

9 8 7 6 5 4 3 2 1

q26.3

q26.2

q26.1

q25.3

q25.2

q25.1

q24.2 q24.3

q24.1

q23

q22.31 q22.32 q22.33

q22.2

q22.1

q21.3

q21.2

q21.1

q15.1 q15.2 q15.3

q14

q13.1 q13.2 q13.3

q12

q11.2

q11.1

p11.1

p11.2

p12

p13


●

PAM50 region31 chr15:96304947−97608945 Subtypes Segregation − from ARRDC4 to TTC23 PAM50

14


12

F

10 8 6 4 2 TTC23

SYNM

IGF1R

FAM169B

ARRDC4

0


IntClust region31 chr15:96304947−97608945 Subtypes Segregation − from ARRDC4 to TTC23 IntClust

14

1 2 3 4 5 6 7 8 9 10

12

F

10 8 6 4 2 TTC23

SYNM

IGF1R

FAM169B

ARRDC4

0


k−means region31 chr15:96304947−97608945 Subtypes Segregation − from ARRDC4 to TTC23 k−means

14

9 8 7 6 5 4 3 2 1

12 8 6 4 2


TTC23

SYNM

IGF1R

FAM169B

0 ARRDC4

F

10

q24.2 q24.3

q24.1

q23.3

q23.2

q23.1

q22.3

q22.2

q22.1

q21

q13

q12.2

q12.1

q11.2

q11.1

p11.1

p11.2

p12.1

p12.2

p12.3

p13.11

p13.12

p13.13

p13.2

p13.3


●

PAM50 region32 chr16:67411784−68661883 Subtypes Segregation − from TMCO7 to PDXDC2 8


F

6 4 2

PDXDC2

CLEC18A

WWP2

NOB1

NQO1

NFAT5

CYB5B

TERF2

TMED6

NIP7

COG8

VPS4A

SNTB2

CIRH1A

CHTF8

HAS3

TMCO7

0


IntClust region32 chr16:67411784−68661883 Subtypes Segregation − from TMCO7 to PDXDC2 8

IntClust 1 2 3 4 5 6 7 8 9 10

F

6 4 2

PDXDC2

CLEC18A

WWP2

NOB1

NQO1

NFAT5

CYB5B

TERF2

TMED6

NIP7

COG8

VPS4A

SNTB2

CIRH1A

CHTF8

HAS3

TMCO7

0


k−means region32 chr16:67411784−68661883 Subtypes Segregation − from TMCO7 to PDXDC2 8

k−means 9 8 7 6 5 4 3 2 1

4 2


PDXDC2

CLEC18A

WWP2

NOB1

NQO1

NFAT5

CYB5B

TERF2

TMED6

NIP7

COG8

VPS4A

SNTB2

CIRH1A

CHTF8

HAS3

0 TMCO7

F

6

q25.3

q25.2

q25.1

q24.3

q24.2

q24.1

q23.3

q23.2

q23.1

q22

q21.33

q21.32

q21.31

q21.2

q21.1

q12

q11.2

p11.1 q11.1

p11.2

p12

p13.1

p13.2

p13.3


●

PAM50 region33 chr17:11004441−12848195 Subtypes Segregation − from DNAH9 to ELAC2 PAM50

20


F

15 10 5

ELAC2

AC005277.1

MYOCD

AC005358.1

MAP2K4

ZNF18

DNAH9

0


IntClust region33 chr17:11004441−12848195 Subtypes Segregation − from DNAH9 to ELAC2 IntClust

20

1 2 3 4 5 6 7 8 9 10

F

15 10 5

ELAC2

AC005277.1

MYOCD

AC005358.1

MAP2K4

ZNF18

DNAH9

0


k−means region33 chr17:11004441−12848195 Subtypes Segregation − from DNAH9 to ELAC2 k−means

20

9 8 7 6 5 4 3 2 1

F

15 10 5


ELAC2

AC005277.1

MYOCD

AC005358.1

MAP2K4

ZNF18

DNAH9

0

USP22

IFT20

TNFAIP1

SARM1


SSH2

GOSR1

CPD

TMIGD1

*

GOSR1

BLMH

SLC6A4

CCDC55

EFCAB5

SSH2

CORO6

ANKRD13B

*

CPD

TMIGD1

BLMH

SLC6A4

CCDC55

* GIT1

TP53I13

AC104564.11−1

TAOK1

*

EFCAB5

* *

CORO6

*

GIT1

* NUFIP2

CRYBA1

*

ANKRD13B

PIPOX MYO18A

*

TP53I13

* SEZ6

* *

AC104564.11−1

* PHF12

DHRS13

*

TAOK1

* *

NUFIP2

*

CRYBA1

*

PIPOX

25 *

MYO18A

* *

SEZ6

*

PHF12

*

DHRS13

*

FLOT2

*

TRAF4

25

ERAL1

GOSR1

CPD

TMIGD1

BLMH

SLC6A4

CCDC55

EFCAB5

SSH2

CORO6

ANKRD13B

GIT1

TP53I13

AC104564.11−1

TAOK1

NUFIP2

CRYBA1

MYO18A

PIPOX

SEZ6

PHF12

DHRS13

FLOT2

ERAL1

TRAF4

NEK8

TLCD1

RAB34

AC005726.8

*

FLOT2

NEK8

SDF2 SUPT6H

*

TRAF4

* TLCD1

RAB34

AC005726.8

SPAG5

25

ERAL1

* SDF2 *

NEK8

* KIAA0100

AC005726.6

*

TLCD1

20

RAB34

* *

SUPT6H

20

AC005726.8

* *

SDF2

*

KIAA0100

ALDOC

PIGS

20

SUPT6H

* SPAG5

FOXN1 UNC119

*

KIAA0100

15 ALDOC

*

AC005726.6

*

SPAG5

*

PIGS

*

ALDOC

FOXN1

*

AC005726.6

10 UNC119

SLC13A2

SLC46A1

SARM1

VTN

SEBOX

C17orf32 *

PIGS

SARM1 SLC13A2

SLC46A1

*

UNC119

FOXN1

SLC13A2

SLC46A1

VTN

SEBOX

C17orf32

*

VTN

* POLDIP2

TNFAIP1

IFT20

TMEM97

PPY2

NLK

AC130289.2

NOS2

LGALS9

WSB1

FAM27L

KCNJ12

MAP2K3

C17orf103

15

SEBOX

USP22 TMEM11

DHRS7B

F

*

C17orf32

IFT20 TNFAIP1 *

POLDIP2

*

POLDIP2

TMEM97

PPY2

NLK

AC130289.2

F 15

TMEM97

PPY2

NLK

NOS2

LGALS9

WSB1

FAM27L

KCNJ12

MAP2K3

C17orf103

*

AC130289.2

NOS2

LGALS9

USP22 TMEM11

DHRS7B

10

WSB1

FAM27L

KCNJ12

MAP2K3

C17orf103

TMEM11

DHRS7B

F

*

*

*

*

5

*

0 q25.3

q25.2

q25.1

q24.3

q24.2

q24.1

q23.3

q23.2

q23.1

q22

q21.33

q21.32

q21.31

q21.2

q21.1

q12

q11.2

p11.1 q11.1

p11.2

p12

p13.1

p13.2

p13.3


●

PAM50 region34 chr17:20843498−25877958 Subtypes Segregation − from USP22 to GOSR1

*

PAM50

*

*


10

5 *

0


IntClust region34 chr17:20843498−25877958 Subtypes Segregation − from USP22 to GOSR1 IntClust

* *

* *

* *

5 1 2 3 4 5 6 7 8 9 10

0


k−means region34 chr17:20843498−25877958 Subtypes Segregation − from USP22 to GOSR1

k−means

9 8 7 6 5 4 3 2 1

q25.3

q25.2

q25.1

q24.3

q24.2

q24.1

q23.3

q23.2

q23.1

q22

q21.33

q21.32

q21.31

q21.2

q21.1

q12

q11.2

p11.1 q11.1

p11.2

p12

p13.1

p13.2

p13.3


●

PAM50 region35 chr17:35015230−35273967 Subtypes Segregation − from NEUROD2 to IKZF3 100

PAM50 *

* *

80


* *

F

60 *

*

40

*

20

IKZF3

GRB7

C17orf37

ERBB2

PERLD1

PNMT

TCAP

STARD3

PPP1R1B

NEUROD2

0


IntClust region35 chr17:35015230−35273967 Subtypes Segregation − from NEUROD2 to IKZF3 100

IntClust *

* *

*

80

*

F

60 *

*

40

* *

20

*

1 2 3 4 5 6 7 8 9 10

IKZF3

GRB7

C17orf37

ERBB2

PERLD1

PNMT

TCAP

STARD3

PPP1R1B

NEUROD2

0


k−means region35 chr17:35015230−35273967 Subtypes Segregation − from NEUROD2 to IKZF3 100

k−means *

* *

*

80

*

F

60 *

*

40

* *

20

*


IKZF3

GRB7

C17orf37

ERBB2

PERLD1

PNMT

TCAP

STARD3

PPP1R1B

NEUROD2

0

9 8 7 6 5 4 3 2 1

TAC4

DLX3

*


*

*

0

*

*

* SPAG9

50

50 NME1−NME2

SPAG9 NME1−NME2

*

SPAG9

*

TOB1

WFIKKN2

50

NME1−NME2

*

TOB1

WFIKKN2

C17orf73

AC005921.3−1

ANKRD40

ABCC3

CACNA1G

SPATA20

EPN3

MYCBPAP

RSAD1

CHAD

LRRC59

*

TOB1

C17orf73

ANKRD40

0 AC005921.3−1

ABCC3

CACNA1G

* *

WFIKKN2

EPN3 SPATA20

* *

C17orf73

*

ANKRD40

*

AC005921.3−1

*

ABCC3

10 *

CACNA1G

*

EPN3

* *

SPATA20

*

RSAD1

*

MYCBPAP

10

CHAD

EME1

*

RSAD1

* LRRC59

XYLT2

*

MYCBPAP

* *

CHAD

* *

LRRC59

*

EME1

* MRPL27

*

EME1

* XYLT2

*

MRPL27

TMEM92

COL1A1

HILS1

SGCA

PPP1R9B

SAMD14

PDK2

*

MRPL27

TMEM92

COL1A1

HILS1

SGCA

PPP1R9B

SAMD14

PDK2

ITGA3

DLX3

DLX4

TAC4

MYST2

FAM117A

SLC35B1

SPOP

NXPH3

*

XYLT2

TMEM92

COL1A1

HILS1

*

SGCA

*

SAMD14

DLX3

*

PPP1R9B

* ITGA3

*

PDK2

TAC4 DLX4 *

ITGA3

* MYST2

*

DLX4

30 FAM117A

PHB NGFR

*

MYST2

*

FAM117A

30

SLC35B1

SPOP

NXPH3

NGFR

ZNF652

AC004797.1

PHOSPHO1

ABI3

GNGT2

B4GALNT2

IGF2BP1

GIP

SNF8

30

SLC35B1

SPOP

NXPH3

40

NGFR

40

PHB

ZNF652

AC004797.1

PHOSPHO1

ABI3

GNGT2

B4GALNT2

IGF2BP1

GIP

F 40

PHB

ZNF652

AC004797.1

PHOSPHO1

ABI3

GNGT2

B4GALNT2

IGF2BP1

* SNF8

0 *

UBE2Z

ATP5G1

10 *

GIP

* UBE2Z

* *

SNF8

* ATP5G1

* *

UBE2Z

20 CALCOCO2

TTLL6

HOXB13

*

ATP5G1

20

CALCOCO2

TTLL6

HOXB13

F 20

CALCOCO2

TTLL6

HOXB13

F 60

60

60

*

*

*

*

*

q25.3

q25.2

q25.1

q24.3

q24.2

q24.1

q23.3

q23.2

q23.1

q22

q21.33

q21.32

q21.31

q21.2

q21.1

q12

q11.2

p11.1 q11.1

p11.2

p12

p13.1

p13.2

p13.3


●

PAM50 region36 chr17:44157125−46621138 Subtypes Segregation − from HOXB13 to NME1−NME2

*

PAM50

* *


* *

*


IntClust region36 chr17:44157125−46621138 Subtypes Segregation − from HOXB13 to NME1−NME2

*

IntClust

* * * *

1 2 3 4 5 6 7 8 9 10


k−means region36 chr17:44157125−46621138 Subtypes Segregation − from HOXB13 to NME1−NME2

k−means

9 8 7 6 5 4 3 2 1

q25.3

q25.2

q25.1

q24.3

q24.2

q24.1

q23.3

q23.2

q23.1

q22

q21.33

q21.32

q21.31

q21.2

q21.1

q12

q11.2

p11.1 q11.1

p11.2

p12

p13.1

p13.2

p13.3


●

PAM50 region37 chr17:53777538−53849930 Subtypes Segregation − from SUPT4H1 to RNF43 40

PAM50 *

F

30

*


20 10

RNF43

SUPT4H1

0


IntClust region37 chr17:53777538−53849930 Subtypes Segregation − from SUPT4H1 to RNF43 40

IntClust *

F

30

*

20 10

1 2 3 4 5 6 7 8 9 10

RNF43

SUPT4H1

0


k−means region37 chr17:53777538−53849930 Subtypes Segregation − from SUPT4H1 to RNF43 40

k−means * *

20 10

RNF43

0 SUPT4H1

F

30


9 8 7 6 5 4 3 2 1

C17orf28 CDR2L ICT1 ATP5H KCTD2 SLC16A5 ARMC7 NT5C AC022211.5−2 NUP85 GGA3 MRPS7 MIF4GD SLC25A19 GRB2 AC011933.2 KIAA0195 CASKIN2 TSEN54 LLGL2 RECQL5 AC087749.12−2 SAP30BP ITGB4 GALK1 H3F3B UNK UNC13D WBP2 TRIM47 TRIM65 MRPL38 FBF1 ACOX1 CDK3 EVPL SRP68 GALR2 ZACN EXOC7 FOXJ1 RNF157 FAM100B QRICH2 PRPSAP1 SPHK1 UBE2O AANAT RHBDF2 CYGB ST6GALNAC2 ST6GALNAC1 MXRA7 JMJD6 C17orf95 SFRS2 MGAT5B SEC14L1 SEPT9 AC015804.14 TNRC6C TMC6 TMC8 C17orf99 SYNGR2 TK1 AFMID BIRC5 AC087645.19−1 SOCS3 PGS1 DNAH17 CYTH1 TIMP2 AC100788.1 LGALS3BP CANT1 C1QTNF1 ENGASE AC021534.1 ENPP7 CBX2 CBX8 CBX4 TBC1D16 GAA EIF4A3 CARD14 SGSH SLC26A11 KIAA1618 RNF213 AC120024.22−2 NPTX1 AC127496.5−2 CHMP6 AC127496.5−3 BAIAP2 AATK AZI1 C17orf56 SLC38A10 C17orf55 TMEM105 ACTG1 FSCN2 C17orf70 NPLOC4 PDE6G C17orf90 CCDC137 ARL16 HGS MRPL12 SLC25A10 DYSFIP1 ARHGDIA ANAPC11 PCYT2 SIRT7 MAFG PYCR1 MYADML2 NOTUM ASPSCR1 STRA13 LRRC45 RAC3 DCXR RFNG GPS1 DUS1L FASN

F 10

5


F


F

*

* * * *

* * * *

**

*

15 *

* *

* *

* * * *

* * * * *

*

* *

* * *

* *

* *

*


15

10

5

0 q25.3

q25.2

q25.1

q24.3

q24.2

q24.1

q23.3

q23.2

q23.1

q22

q21.33

q21.32

q21.31

q21.2

q21.1

q12

q11.2

p11.1 q11.1

p11.2

p12

p13.1

p13.2

p13.3


●

PAM50 region38 chr17:70458434−77649395 Subtypes Segregation − from C17orf28 to FASN

15 *

PAM50


5

0


IntClust region38 chr17:70458434−77649395 Subtypes Segregation − from C17orf28 to FASN

*

IntClust

* * **

* *

* **

* *

1 2 3 4 5 6 7 8 9 10

0


k−means region38 chr17:70458434−77649395 Subtypes Segregation − from C17orf28 to FASN

k−means

9 8 7 6 5 4 3 2 1

q13.43

q13.42

q13.41

q13.33

q13.32

q13.31

q13.2

q13.13

q13.12

q13.11

q12

q11

p11

p12

p13.11

p13.12

p13.13

p13.2

p13.3


●

PAM50 region39 chr19:15131444−15351603 Subtypes Segregation − from NOTCH3 to AKAP8 10


8

F

6 4 2

AKAP8

BRD4

ABHD9

NOTCH3

0


IntClust region39 chr19:15131444−15351603 Subtypes Segregation − from NOTCH3 to AKAP8 10

IntClust 1 2 3 4 5 6 7 8 9 10

8

F

6 4 2

AKAP8

BRD4

ABHD9

NOTCH3

0


k−means region39 chr19:15131444−15351603 Subtypes Segregation − from NOTCH3 to AKAP8 10

k−means 9 8 7 6 5 4 3 2 1

8

F

6 4 2


AKAP8

BRD4

ABHD9

NOTCH3

0


ATF5

NUP62

IL4I1

TBC1D17

AKT1S1

PNKP

PTOV1

MED25

FUZ

AP2A1

TSKS

CPT1C

PRMT1

BCL2L12

IRF3

SCAF1

RRAS

PRR12

PRRG2

NOSIP

RCN3

FCGRT

25

RPL13A

FLT3LG

ALDH16A1

PIH1D1

F

ATF5

NUP62

IL4I1

TBC1D17

AKT1S1

PNKP

PTOV1

MED25

FUZ

AP2A1

TSKS

CPT1C

PRMT1

BCL2L12

IRF3

SCAF1

RRAS

PRR12

PRRG2

NOSIP

RCN3

FCGRT

RPL13A

FLT3LG

ALDH16A1

PIH1D1

F

ATF5

NUP62

IL4I1

TBC1D17

AKT1S1

PNKP

PTOV1

MED25

FUZ

AP2A1

TSKS

CPT1C

PRMT1

BCL2L12

IRF3

SCAF1

RRAS

PRR12

PRRG2

NOSIP

RCN3

FCGRT

RPL13A

FLT3LG

ALDH16A1

PIH1D1

F

*

20

15

10

5

0

25

20

15

25

20

15

10

5 q13.43

q13.42

q13.41

q13.33

q13.32

q13.31

q13.2

q13.13

q13.12

q13.11

q12

q11

p11

p12

p13.11

p13.12

p13.13

p13.2

p13.3


●

PAM50 region40 chr19:54641368−55129004 Subtypes Segregation − from PIH1D1 to ATF5 PAM50 Normal LumB LumA Her2 Basal

10

5

0


IntClust region40 chr19:54641368−55129004 Subtypes Segregation − from PIH1D1 to ATF5 IntClust 1 2 3 4 5 6 7 8 9 10

0


k−means region40 chr19:54641368−55129004 Subtypes Segregation − from PIH1D1 to ATF5

k−means

9 8 7 6 5 4 3 2 1

CTNNBL1


SYS1

SDC4

RBPSUHL

MATN4

SEMG2

SEMG1

PI3

WFDC12

WFDC5

KCNS1

STK4

TOMM34

YWHAB

RIMS4

KCNK15

WISP2

ADA

PKIG

SERINC3

TTPAL

C20orf62

HNF4A

R3HDML

C20orf142

GDAP1L1

C20orf111

JPH2

C20orf100

FAM112A

MYBL2

IFT52

SGK2

L3MBTL

SFRS6

PTPRT

SYS1

SDC4

RBPSUHL

MATN4

SEMG2

SEMG1

PI3

WFDC12

WFDC5

KCNS1

STK4

TOMM34

YWHAB

RIMS4

KCNK15

WISP2

ADA

PKIG

SERINC3

TTPAL

C20orf62

HNF4A

R3HDML

C20orf142

GDAP1L1

C20orf111

JPH2

C20orf100

FAM112A

MYBL2

IFT52

SGK2

L3MBTL

SFRS6

PTPRT

CHD6

EMILIN3

LPIN3

ZHX3

PLCG1

TOP1

MAFB

DHX35

FAM83D

PPP1R16B

ACTR5

SLC32A1

ADIG

KIAA1219

SNHG11

LBP

BPI

KIAA1755

TGM2

KIAA0406

F 8

CHD6

EMILIN3

LPIN3

ZHX3

CTNNBL1 C20orf102

6

PLCG1

TOP1

MAFB

DHX35

FAM83D

PPP1R16B

ACTR5

SLC32A1

ADIG

KIAA1219

SNHG11

LBP

BPI

KIAA1755

TGM2

KIAA0406

12

C20orf102

F

CTNNBL1

SYS1

SDC4

RBPSUHL

MATN4

SEMG2

SEMG1

PI3

WFDC12

WFDC5

KCNS1

STK4

TOMM34

YWHAB

RIMS4

KCNK15

WISP2

ADA

PKIG

SERINC3

TTPAL

C20orf62

HNF4A

R3HDML

C20orf142

GDAP1L1

C20orf111

JPH2

C20orf100

FAM112A

MYBL2

IFT52

SGK2

L3MBTL

SFRS6

PTPRT

CHD6

EMILIN3

LPIN3

ZHX3

PLCG1

TOP1

MAFB

DHX35

FAM83D

PPP1R16B

ACTR5

SLC32A1

ADIG

KIAA1219

SNHG11

LBP

BPI

KIAA1755

TGM2

KIAA0406

C20orf102

F

14

*

10

8

6

4

2

0 q13.33

q13.32

q13.31

q13.2

q13.13

q13.12

q13.11

q12

q11.23

q11.22

q11.21

q11.1

p11.1

p11.21

p11.22

p11.23

p12.1

p12.2

p12.3

p13


●

PAM50 region41 chr20:35755848−43440371 Subtypes Segregation − from CTNNBL1 to SYS1

14 PAM50

12

10


6

4

2

0


IntClust region41 chr20:35755848−43440371 Subtypes Segregation − from CTNNBL1 to SYS1

14 IntClust

12

10

* *

4

2 1 2 3 4 5 6 7 8 9 10

0


k−means region41 chr20:35755848−43440371 Subtypes Segregation − from CTNNBL1 to SYS1

k−means

9 8 7 6 5 4 3 2 1

TP53RK

SPATA2

KCNG1


C20orf43

SPO11

BMP7

TFAP2C

C20orf107

*

C20orf106

*

CASS4

*

CSTF1

SPO11

BMP7

TFAP2C

C20orf107

C20orf106

C20orf43

CASS4

CSTF1

AURKA

C20orf108

MC3R

CBLN4

DOK5

PFDN4

CYP24A1

BCAS1

ZNF217

TSHZ2

ZFP64

SALL4

ATP9A

NFATC2

KCNG1

MOCS3

DPM1

ADNP

BCAS4

PARD6B

C20orf175

30

AURKA

C20orf108

MC3R

CBLN4

DOK5

PFDN4

CYP24A1

BCAS1

30

ZNF217

TSHZ2

*

ZFP64

SALL4

ATP9A

NFATC2

*

*

MOCS3

*

DPM1

*

ADNP

*

BCAS4

PTPN1

AL354889.14

CEBPB

UBE2V1

*

PARD6B

C20orf175

*

PTPN1

AL354889.14

40

CEBPB

* SNAI1

F 40

UBE2V1

* RNF114

SPATA2

SLC9A8

B4GALT5

PTGIS

KCNB1

ZNFX1

DDX27

STAU1

CSE1L

ARFGEF2

AL035106.28

PREX1

SULF2

NCOA3

ZMYND8

EYA2

*

SNAI1

TP53RK SLC2A10

20

RNF114

*

SLC9A8

B4GALT5

PTGIS

KCNB1

*

ZNFX1

*

STAU1

CSE1L

*

DDX27

20

ARFGEF2

AL035106.28

PREX1

SULF2

NCOA3

ZMYND8

EYA2

10

SLC2A10

F

TP53RK

SPO11

BMP7

TFAP2C

C20orf107

C20orf106

C20orf43

CASS4

CSTF1

AURKA

C20orf108

MC3R

CBLN4

DOK5

PFDN4

CYP24A1

BCAS1

ZNF217

TSHZ2

ZFP64

SALL4

ATP9A

NFATC2

KCNG1

MOCS3

DPM1

ADNP

BCAS4

PARD6B

C20orf175

PTPN1

AL354889.14

CEBPB

UBE2V1

SNAI1

RNF114

SPATA2

SLC9A8

B4GALT5

PTGIS

KCNB1

ZNFX1

DDX27

STAU1

CSE1L

ARFGEF2

AL035106.28

PREX1

SULF2

NCOA3

ZMYND8

EYA2

SLC2A10

F

*

*

*

*

*

*

*

0 q13.33

q13.32

q13.31

q13.2

q13.13

q13.12

q13.11

q12

q11.23

q11.22

q11.21

q11.1

p11.1

p11.21

p11.22

p11.23

p12.1

p12.2

p12.3

p13


●

PAM50 region42 chr20:44746411−55386926 Subtypes Segregation − from TP53RK to SPO11

40 PAM50


20

10

0


IntClust region42 chr20:44746411−55386926 Subtypes Segregation − from TP53RK to SPO11

*

IntClust

* *

*

10 *

1 2 3 4 5 6 7 8 9 10

0


k−means region42 chr20:44746411−55386926 Subtypes Segregation − from TP53RK to SPO11

k−means

9 8 7 6 5 4 3 2 1

0

Genes & Chi−square Test C20orf201

OPRL1

RGS19

TCEA2

SOX18

PRR17

PRPF6

SAMD10

ZNF512B

UCKL1

DNAJC5

C20orf135

TPD52L2

ZBTB46

SLC2A4RG

LIME1

ZGPAT

ARFRP1

5

C20orf201

OPRL1

RGS19

TCEA2

SOX18

PRR17

PRPF6

*

SAMD10

5

*

ZNF512B

*

*

UCKL1

DNAJC5

C20orf135

TPD52L2

ZBTB46

SLC2A4RG

LIME1

ZGPAT

ARFRP1

15 RTEL1

STMN3

GMEB2

RP4−697K14.11

C20orf195

SRMS

PTK6

F 15

RTEL1

STMN3

GMEB2

RP4−697K14.11

C20orf195

10

SRMS

PTK6

F

C20orf201

OPRL1

RGS19

TCEA2

SOX18

PRR17

PRPF6

SAMD10

ZNF512B

UCKL1

DNAJC5

C20orf135

TPD52L2

ZBTB46

SLC2A4RG

LIME1

ZGPAT

ARFRP1

RTEL1

STMN3

GMEB2

RP4−697K14.11

C20orf195

SRMS

PTK6

F

*

q13.33

q13.32

q13.31

q13.2

q13.13

q13.12

q13.11

q12

q11.23

q11.22

q11.21

q11.1

p11.1

p11.21

p11.22

p11.23

p12.1

p12.2

p12.3

p13


●

PAM50 region43 chr20:61622577−62186156 Subtypes Segregation − from PTK6 to C20orf201

15 PAM50


5

0


IntClust region43 chr20:61622577−62186156 Subtypes Segregation − from PTK6 to C20orf201

*

IntClust

10 *

*

1 2 3 4 5 6 7 8 9 10

0


k−means region43 chr20:61622577−62186156 Subtypes Segregation − from PTK6 to C20orf201

k−means

9 8 7 6 5 4 3 2 1

q22.3

q22.2

q22.13

q22.12

q22.11

q21.3

q21.2

q21.1

q11.2

q11.1

p11.1

p11.2

p12

p13


●

PAM50 region44 chr21:46479476−46690110 Subtypes Segregation − from C21orf57 to PCNT 10


8

F

6 4 2

PCNT

C21orf58

C21orf57

0


IntClust region44 chr21:46479476−46690110 Subtypes Segregation − from C21orf57 to PCNT 10

IntClust *

8

F

6 4 2

1 2 3 4 5 6 7 8 9 10

PCNT

C21orf58

C21orf57

0


k−means region44 chr21:46479476−46690110 Subtypes Segregation − from C21orf57 to PCNT 10

k−means 9 8 7 6 5 4 3 2 1

8

F

6 4 2

PCNT

C21orf58

C21orf57

0


q28

q27.3

q27.1 q27.2

q26.3

q26.1 q26.2

q25

q24

q23

q22.3

q22.1 q22.2

q21.33

q21.32

q21.31

q21.2

q21.1

q13.2 q13.3

q13.1

q12

p11.21 p11.1 q11.1 q11.2

p11.22

p11.23

p11.3

p11.4

p21.1

p21.2

p21.3

p22.13 p22.12 p22.11

p22.2

p22.33 p22.32 p22.31

Chromosome X deletion outliers

●

PAM50 region45 chrX:1482032−2429015 Subtypes Segregation − from ASMTL to ZBED1 PAM50

8


F

6 4 2

ZBED1

DHRSX

ASMT

CXYorf3

P2RY8

ASMTL

0


IntClust region45 chrX:1482032−2429015 Subtypes Segregation − from ASMTL to ZBED1 IntClust

8

1 2 3 4 5 6 7 8 9 10

F

6 4 2

ZBED1

DHRSX

ASMT

CXYorf3

P2RY8

ASMTL

0


k−means region45 chrX:1482032−2429015 Subtypes Segregation − from ASMTL to ZBED1 k−means

8

9 8 7 6 5 4 3 2 1

F

6 4 2


ZBED1

DHRSX

ASMT

CXYorf3

P2RY8

ASMTL

0

Figure S15: Patterns of CNA driven (HMM-based) outlying expression at selected regions throughout the genome. The top track shows the region of interest (red box) on a chromosome ideogram. The three sets of barplots indicate the subtype distribution for PAM50, IntClust, and K-Means derived subgroups for each gene found within the region of interest. Red asterisks above the bar plots indicate significantly different observed distributions than expected based on the overall population frequency (χ2 , P-value < 0.0001).

100

1 2 3 4 5 6 7 8 9 10

1 2 3 4 5 6 7 8 9 10

1 2 3 4 5 6 7 8 9 10

1 2 3 4 5 6 7 8 9 10

1 2 3 4 5 6 7 8 9 10

1 2 3 4 5 6 7 8 9 10

1 2 3 4 5 6 7 8 9 10

Gene Expression LogR 3

2

1

0

−1 Gene Expression LogR

4 4

2

0

−2 3

2

1

−1 0 −2

−4

5 5

4 4

3

2

1

0

−1

−2

101

1 2 3 4 5 6 7 8 9 10

0

−2

6

4

2

0

−2

−4

4

2

0

−2

6 4

2

0

−2

Gene Expression LogR

LumB

IGF1R

6

6

4

2

0

−2

−4

3

2

1

0

−1

−2

−3

Her2

ERBB2

LumB Normal

−4 LumA





−2

Basal

Normal

LumB

LumA

Her2

Basal

Normal

LumB

LumA

Her2

Basal

Normal

LumB

LumA

Her2

Basal

Normal

0

Normal

FOXM1

c11orf67

LumB

0

2

2

Her2

2

6

LumA

−4 4

4 Gene Expression LogR

6

4

Basal

−2

−4


6

−2


1 2 3 4 5 6 7 8 9 10

−2

0

LumB

ZNF703

INTS4

Normal

0 0

2

Her2

2

6

LumA

4

2

4

Basal

6

4 Gene Expression LogR

Her2

−4


−2

−2


1 2 3 4 5 6 7 8 9 10

0

0

LumB

−4

PAK1

Normal

−2 2

2

Her2

0

6

LumA

2 4


4


ZMIZ1

LumA

−4

4

Basal

4 −2


6

0

LumB

−4 2

Normal

−2 Basal

CLNS1A

LumA

1 2 3 4 5 6 7 8 9 10

0


1 2 3 4 5 6 7 8 9 10


Her2

−4 2


−4

4

Basal

−2 −2


0 0

LumB

2 2

Normal

AIM2 LumB

c11orf30

LumA

6 Normal

6

Her2

−3 4

Basal

−2


0


−1

LumB

1

Normal

−4 Her2

−2

LumA

0

Basal

2


CCND1

LumA

1 2 3 4 5 6 7 8 9 10

2


Normal

LumB

LumA

Her2

Basal


Her2

4


1 2 3 4 5 6 7 8 9 10


Basal

Normal

LumB

LumA

Her2

Basal


a 6

RSF1

4

2

−2

0

−4

6

4

2

−2

0

Int Clust Group Int Clust Group Int Clust Group Int Clust Group Int Clust Group Int Clust Group Int Clust Group

6

MDM2

4

2

−2

0

−4

8

6

4

2

−2

0

Int Clust Group Int Clust Group Int Clust Group Int Clust Group Int Clust Group Int Clust Group Int Clust Group

b

8

9 10

6

8

9 10

6 4 2 0


−4

1

3

5

7

Her2

LumA

LumB Normal

1

3

5

7

4 2 −4

−4

4

Int Clust Group

Basal

6

LumB Normal



4 2

2

−2

2

LumA

0

7

Her2

−2

5

0


6

4

6

3

−4

4

Int Clust Group

0


−4

1

−2

4 2 0 −2

2

Basal

6

7

LumB Normal

4

5

LumA

2

3

Her2

0

1

Basal

−2

LumB Normal

6

LumA

MTAP

−2

4 2 0


−4

−2

4 2 0


−2 −4

Her2

6

Basal

−4


MAP2K4

6

PPP2R2

6

PTEN

2

4

6

Int Clust Group

8

9 10

2

4

6

8

9 10

Int Clust Group

Figure S16: Boxplots of log2 expression stratified by the PAM50 or integrative subtypes are shown for select genomic regions. As noted in the text, these regions harbour a. predicted amplifications or b. homozygous deletions.

102

Figure S17: Probe-level plots of putative homozygous deletions referred to in the text. a. Putative homozygous deletions as predicted by CBS analysis are shown for PTEN, where red indicates regions of gain, black indicates copy number neutral regions, and blue indicates regions of loss. b. As in (a) for PPP2R2A. c. As in (a) for MTAP. d. As in (a) for MAP2K4. 106

Figure S18: Putative MTAP homozygous deletions as predicted by HMM analysis. Here red indicates regions of gain, blue indicates copy number neutral regions, and green indicates regions of loss, where more intense shading corresponds to higher amplitude events. 107

Figure S19: Putative MAP2K4 homozygous deletions as predicted by HMM analysis. Here red indicates region of gain, blue indicates copy number neutral regions, and green indicates regions of loss, where more intense shading corresponds to higher amplitude events. 108

TRA

G TR

TNFRSF4

b

TNFRSF13B

TNFRSF9 TNFSF13B

TNFRSF17 CD27

TRAF3

STK4

TNFAIP3

CYTH1 CD226 RASSF5

TANK

SELE

ETS1

TRAF1

TNFSF14

THBS3

GZMK

LYZ VCAM1

ITGA4

GZMH CYFIP2 TAGAP ARHGAP15 ARHGAP25

AMICA1

ICOS

SELL

CD40LG

CRLF3

SH3BP1 RNASE6 CD48

HCST

KLRG1

LTB

CORO1A

LCP1

ITGAX SP110

GZMA

IL16

RHOF ITGB2

FCER2

PIK3CD INPP5D

GPSM3

RASSF2

RAC2

FASLG

MATK

DGKA

NOTCH3

PLCG2 TNFAIP8

SH2D2A PTPRC

FYB LAT

PAG1

TAP1

IL21R

NFATC3

CD5

TYROBP

PTPN7

CD8A

ZAP70

HLA-DRA

PTPN2

CD6

IRF9

PSMB10 APOBEC3G

IFNG

SRGN

PSMB9

HLA-DPA1 CD3E

HLA-DRB3

PSME1

STAT4

CD3D

ISG20

STAT1

SOCS1

ITK TRAF3IP3

SH2D1A

IL3RA S1PR4 FAIM3

IL2RG

MAPK1 GRAP2

FYN

PRF1

MYO1F

GRAP VAV1

LCP2 SLAMF1 TRAT1

CD52

GZMB

CD2

IL7R

LYN

NCK1

PTPRCAP

GAB3

TXK

KLRB1

WIPF1

ACAP1

CD19 SEMA4D

0 0 0 0 0 0 0

LTA

ACTR2 SP100

IL2RA

CD8B

IL18R1

IL15 HLA-A IL12RB1

HLA-DPB1 B2M

PSMB8

HLA-F HLA-G

SIT1 PTPN22

NCR3

IL32

SLA2 CD74 CD1B

KLRD1

KLRC3

BTN3A3

CD247

KLRC1

KIR2DL4 KIR3DL1 CD97

CCL5

Module 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

CXCL9

CXCR3 CXCL13

CXCR5 CCL19 WNT10A

GNG2

CCR7

GPR18

PLCB2 CXCR6 CCL21

CCR4

CCL22

P2RY10

0 0 0 0 0 0 0 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 3

CCL17

3 3 3 3 3 3 4 4 4 4 4 5 5 5 6 7 7 7 7 7 7 7 7 7 7 7

GeneSet TCR signaling(R) TCR signaling in naive CD4+ T cells(N) T cell receptor signaling pathway(K) TCR signaling in naive CD8+ T cells(N) T cell activation(P) Primary immunodeficiency(K) Hematopoietic cell lineage(K) Fc-epsilon receptor I signaling in mast cells(N) Antigen processing and presentation(K) t cell receptor signaling pathway(B) role of mef2d in t-cell apoptosis(B) Natural killer cell mediated cytotoxicity(K) lck and fyn tyrosine kinases in initiation of tcr activation(B) Class I PI3K signaling events(N) activation of csk by camp-dependent protein kinase inhibits signaling through the t cell receptor(B) Cell adhesion molecules (CAMs)(K) the co-stimulatory signal during t-cell activation(B) TRAIL signaling pathway(N) IL12-mediated signaling events(N) IL12 signaling mediated by STAT4(N) Downstream signaling in naive CD8+ T cells(N) Immunoregulatory interactions between a Lymphoid and a non-Lymphoid cell(R) Asthma(K) Viral myocarditis(K) Fc epsilon RI signaling pathway(K) Allograft rejection(K) Graft-versus-host disease(K) Type I diabetes mellitus(K) Intestinal immune network for IgA production(K) Antigen processing and presentation(K) Immunoregulatory interactions between a Lymphoid and a non-Lymphoid cell(R) ras-independent pathway in nk cell-mediated cytotoxicity(B) Natural killer cell mediated cytotoxicity(K) Graft-versus-host disease(K) IL12-mediated signaling events(N) Downstream signaling in naive CD8+ T cells(N) Jak-STAT signaling pathway(K) IL12-mediated signaling events(N) Downstream signaling in naive CD8+ T cells(N) Natural killer cell mediated cytotoxicity(K) TCR signaling in naive CD8+ T cells(N) IL12 signaling mediated by STAT4(N) TNF receptor signaling pathway(N) EGF receptor signaling pathway(P) granzyme a mediated apoptosis pathway(B) il12 and stat4 dependent signaling pathway in th1 development(B) TRAIL signaling pathway(N) Inflammation mediated by chemokine and cytokine signaling pathway(P) IFN-gamma pathway(N) IL23-mediated signaling events(N) VEGF signaling pathway(K) IL27-mediated signaling events(N) Chemokine signaling pathway(K) Cytokine-cytokine receptor interaction(K) Receptor-ligand complexes bind G proteins(R) Class A/1 (Rhodopsin-like receptors)(R) Inflammation mediated by chemokine and cytokine signaling pathway(P) Integrin cell surface interactions(R) Immunoregulatory interactions between a Lymphoid and a non-Lymphoid cell(R) Cell surface interactions at the vascular wall(R) Cytokine-cytokine receptor interaction(K) Host Interactions of HIV factors(R) Proteasome(K) Signaling by Wnt(R) M/G1 Transition(R) APC/C-mediated degradation of cell cycle proteins(R) Metabolism of amino acids(R) DNA Replication(R) G1/S Transition(R) S Phase(R) Cell Cycle Checkpoints(R) Apoptosis(R)

c

IL12 SIGNALING MEDIATED BY STAT4(N)

T CELL ACTIVATION(P)

AUTOIMMUNE THYROID DISEASE(K)

ANTIGEN IL2-MEDIATED IL23-MEDIATED INTESTINAL TCR SIGNALING IL12-MEDIATED PROCESSING SIGNALING IMMUNE SIGNALING(R) EVENTS(N) AND EVENTS(N) SIGNALING NETWORK FC-EPSILON PRESENTATION(K) EVENTS(N) FOR IGA T CELL RECEPTOR I CLASS I PI3K PRODUCTION(K) RECEPTOR SIGNALING IN SIGNALING THE SIGNALING RAS-INDEPENDENT MAST EVENTS(N) CO-STIMULATORY GRAFT-VERSUS-HOST PATHWAY(B) PATHWAY IN NK CELLS(N) TYPE I DIABETES SIGNAL DURING LCK AND FYN DISEASE(K) CELL-MEDIATED MELLITUS(K) TYROSINE T-CELL CYTOTOXICITY(B) BCR SIGNALING KINASES ACTIVATION(B) ALLOGRAFT PATHWAY(N) IN INITIATION REJECTION(K) OF RAS SIGNALING IN TCR TCR SIGNALING ACTIVATION(B) CELL ADHESION LEISHMANIASIS(K) IN ROLE OF MEF2D MOLECULES THE CD4+ TCR IN (CAMS)(K) PATHWAY(N) NAIVE CD4+ T TRAIL T-CELL B CELL SIGNALING APOPTOSIS(B) JNK SIGNALING CELLS(N) VIRAL ACTIVATION(P) PATHWAY(N) IN MYOCARDITIS(K) NATURAL CELL SURFACE KILLER HEMATOPOIETIC THE CD4+ TCR INTERACTIONS CELL MEDIATED IL4-MEDIATED CELL LINEAGE(K) PATHWAY(N) AT CYTOTOXICITY(K) SIGNALING THE VASCULAR EVENTS(N) B CELL T CELL WALL(R) RECEPTOR RECEPTOR JAK-STAT PRIMARY SIGNALINGTCR SIGNALING SIGNALING IN SIGNALING IMMUNODEFICIENCY(K) PATHWAY(K) PATHWAY(K) NAIVE CD8+ T PATHWAY(K) CELLS(N) INFLAMMATION IMMUNOREGULATORY CYTOKINE-CYTOKINE MEDIATED BY INTERACTIONS RECEPTOR CHEMOKINE BETWEEN A INTERACTION(K) AND KITRECEPTOR(C) FC EPSILON RI LYMPHOID AND CYTOKINE SIGNALING DOWNSTREAM A SIGNALING PATHWAY(K) CHAGAS SIGNALING IN NON-LYMPHOID CHEMOKINE PATHWAY(P) NAIVE CD8+ T CELL(R) SIGNALING DISEASE(K) CELLS(N) PATHWAY(K)

d CD86

LILRB4

CD80

HLA-DMB

CD40

CD3G

CD74

HLA-DMA

HLA-DOB

CD79B

TNFRSF1B

CD4

CSF2RA

PTPN6

TAPBP

BLK

LCK

CD7

TNF

CD72

LAT2

BTK

SPI1

SYK PTK2B

FCER1G

HMHA1

IL10RA

PARVG

ITGB7

BCAR1

CXCR4

PLEK

FGR HCLS1 WAS

SRGN GNGT2

CECR1

SLA

PNOC

LAPTM5

EVI2B

ELMO1 RGS19

CXCL10

DOCK2

CCR6

LST1

NCF4

CD163

Module 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 5

CST7

GeneSet Alpha-synuclein signaling(N) Fc-epsilon receptor I signaling in mast cells(N) Osteopontin-mediated events(N) EphrinA-EPHA pathway(N) T cell activation(P) TCR signaling(R) Antigen processing and presentation(K) Allograft rejection(K) Graft-versus-host disease(K) Type I diabetes mellitus(K) Intestinal immune network for IgA production(K) Autoimmune thyroid disease(K) Cell adhesion molecules (CAMs)(K) Viral myocarditis(K) IL12 signaling mediated by STAT4(N) IL12-mediated signaling events(N) TCR signaling in naive CD4+ T cells(N) Systemic lupus erythematosus(K) the co-stimulatory signal during t-cell activation(B) Asthma(K) TCR signaling in naive CD8+ T cells(N) activation of csk by camp-dependent protein kinase inhibits signaling through the t cell receptor(B) Receptor-ligand complexes bind G proteins(R)

Figure S20: The T-cell receptor loci on chromosomes 7 and 14 represent trans-acting aberration ‘hotspots’ that modulate an immune response network. a. Circos plot illustrating the connectivity between transacting deletion hotspots at the TCR loci on chromosomes 7 and 14 with their cognate mRNAs. Genes that are significantly associated (P-value < 10−4 ) with deletion of the TRG locus are indicated by purple links, whereas those associated with deletion of TRA are indicated by blue links, and those genes dually regulated by both loci are shown in black. b. Network representation of mRNAs correlated TRG deletion. Coloured subnetworks represent clustering of the network into modules based on an edge-betweenness algorithm (see Methods for additional details). Pathways enriched (FDR < 0.001) for genes clustered in each module are shown on the right. c. Enrichment maps of immune response modules in the trans-associated TRG network. d. As in (b), but for the TRA locus.

112

0.5 0 -0.5

1

2

3

4

5

6

7

8

15 17 19 21 X 9 10 11 12 13 14 16 1 8 20 22

Gene probe location

Frequency

a

Num mRNAs

800

0 1

2

3

4

b CELL CYCLE(K)

5

6

7

8

9

10 11 12 13

CNA location

CDK REGULATION OF G1/S DNA TRANSITION(R) REPLICATION(B)

14

15

17 19 21 X 16 1 8 20 22

SIGNALING BY AURORA KINASES(N)

FOXM1 TRANSCRIPTION FACTOR NETWORK(N)

AURORA B SIGNALING(N)

Chr 5, Basal cancer specific

Frequency

Gene probe location

c

CNA location

Chromosome

d

TRPV6

YES1

EPHB6

DTX3

PTPN4

CASP2

NOTCH1

GFRA1

APH1B ADAM17

NCK1

GTF3C2

MSH6

NRAS LRP8

MMP1 BLM

ERBB3 ERCC1

POLR2FCD2BP2 NR2C2 POLR2D NUP205 YBX1 SNRPA1 POU4F1 XBP1

UPF3B

SF3B3 NFIB TXNL4A

UPF2

GMPS

SFRS5

TFDP1 CSDA

CDH3

RAD54L

TGFA

XAB2 XPC

APBB2

NCK2

ITPR1

BCL2

FOXM1 E2F3 E2F2

MMP7

CDKN2A CDK6 CHEK1

MASTL

MELK

TPX2

TCF7L1

TTK CCNE1 ASPM SLC7A5 SRPK1 CDK4 DLGAP5 CKS1B ESR1 NEK11 AURKA NUP93 CCNA2 CDC45L ESRRA ORC1L CDC25A BIRC5 CDT1 CCNB2 CENPO PLK1 CENPA CAD FOXA1 CENPL MCM6CDK8 ORC6L NCAPD2 KIF2C TRRAP

CEBPB

ACTL6A

MCM10 MCM4

SGOL1 CDCA8 CDC20 AURKB BUB1 RANGAP1 CEP55 RCC2 CENPE

UBE2C

SRF

BTRC

MYB

MIA IL8

HDAC2

GATAD2A IFNAR2DEK UBE3A

DKC1

UBQLN1

TMEM123

PRKAB1 RANBP1

NUTF2 EZH2

LPAR2

BACE2

PCSK4

EPHB1 LMNB2

WDR12 BLOC1S1

FAT1

Module 0 0 1 1 2 2 2 2 3 3 3 3 3 4 4 4 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 6 6 6 7

GeneSet Cell cycle(K) FOXM1 transcription factor network(N) FOXA1 transcription factor network(N) FOXA transcription factor networks(N) M Phase(R) Aurora B signaling(N) Signaling by Aurora kinases(N) mechanism of protein import into the nucleus(B) Formation and Maturation of mRNA Transcript(R) Influenza Life Cycle(R) DNA Repair(R) Transport of Mature Transcript to Cytoplasm(R) Transcription(R) EphrinB-EPHB pathway(N) ErbB signaling pathway(K) Axon guidance(K) G1/S Transition(R) Cell cycle(K) DNA Replication(R) cdk regulation of dna replication(B) M/G1 Transition(R) S Phase(R) Cell Cycle Checkpoints(R) Regulation of retinoblastoma protein(N) G1 Phase(R) Non-small cell lung cancer(K) Glioma(K) cyclins and cell cycle regulation(B) Pancreatic cancer(K) Melanoma(K) Chronic myeloid leukemia(K) cell cycle: g1/s check point(B) Small cell lung cancer(K) FOXM1 transcription factor network(N) Bladder cancer(K) Alzheimer disease-presenilin pathway(P) NOTCH(C) Alzheimer disease-amyloid secretase pathway(P) Regulation of Telomerase(N)

Figure S21: Copy number-expression correlation matrices for CBS and HMM, and the chromosome 5q deletion network module. a. Frequency of genomic copy number aberrations (HMM-derived) (upper panel) and corresponding matrix of copy number and gene expression correlations > 0.3 (middle panel). The frequency of mRNAs associated with a particular copy number aberration is also indicated (lower panel). b. Enrichment maps of cell-cycle and DNA damage response modules for the chromosome 5 deletion transassociated genes. c. The matrix of copy number (CBS-derived) and gene expression profiles with correlation > 0.2 (upper panel) and frequency of copy number events across the entire cohort (lower panel). d. Network representation of mRNAs correlated with deletion of chromosome 5q. Coloured subnetworks represent clustering of the network into modules based on an edge-betweenness algorithm (see Methods for additional details). Pathways enriched (FDR < 0.001) for genes clustered in each module are shown on the right.

115

Figure S22: Joint clustering of copy number and expression data for the top 1,000 cis-acting copy numberexpression associations for CBS-derived CNAs in the discovery cohort (n = 997). Heatmap representation of the joint data matrix for the k =10 clustering, where the product of scaled gene expression and copy number values are shown for the union set of selected features. Genomic instability (GI) based on the proportion of genome altered (black line) and jump measure (red line) is indicated. NPI refers to Nottingham prognostic index.

116

Figure S23: Internal evaluation of cluster validity. a. Dunn’s index was computed for the complete set of features used as input for integrative clustering using the Euclidean distance metric. Evaluation of the indices for k = 2, . . . , 15 indicated a global maximum at k = 14 (ignoring the k = 2 peak) and a local maximum at k = 10 b. The adjusted Rand index (ARI) was computed for all pairwise comparisons of k = 2, . . . , 18 as illustrated in this heatmap, where the value of the ARI is as indicated in the legend. c. Cross tabulation matrices for comparison of k = 9, k = 13, and k = 14 with k = 10.

118

Figure S24: Joint clustering of copy number and expression data for the top 1,000 cis-acting copy numberexpression associations for HMM-derived CNAs in the discovery cohort (n = 997). a. Heatmap representation of the joint data matrix of the union set of selected gene expression and copy number features for the k =10 clustering. b. Kaplan-Meier plots of breast cancer specific survival for the integrative cluster subgroups. c. Kaplan-Meier plots of overall survival for the integrative cluster subgroups.

119

Figure S25: K-Means clustering on outlier gene expression profiles. a. Heatmap representation of K-Means clustering (k = 9) on outlier expression profiles. b. Kaplan-Meier plots of disease-specific survival for the K-Means clusters. c. Kaplan-Meier plots of overall survival for the K-Means clusters.

120

Figure S26: Unsupervised clustering of GI based on two metrics, namely the proportion of genome altered (area), and the ratio between the mean change in log ratio due to alterations versus relative variability (jump) reveals 9 clusters, some of which associate with outcome. a. Unsupervised clustering on GI (derived from CBS) using the Euclidean distance metric and Ward’s method indicates 9 clusters as illustrated in this heatmap representation of the data. b. Graphical representation of the Cox proportional hazard ratios for disease-specific survival for several key variables including, grade, tumour size, age at diagnosis, number of lymph nodes positive, jump, and ER status. The hazard ratio is illustrated for selected values of the covariates, as indicated, where each subgroup was compared against Group 1. Confidence levels correspond to those indicated in the legend. c. Kaplan-Meier plots of disease-specific survival for the GI cluster subgroups. d. Graphical representation of the Cox proportional hazard ratios for overall survival for several key variables including, grade, tumour size, age at diagnosis, number of lymph nodes positive, jump, and ER status. The hazard ratio is illustrated for selected values of the covariates, as indicated, where each subgroup was compared against Group 1. Confidence levels correspond to those indicated in the legend. e. Kaplan-Meier plots of overall survival for the GI cluster subgroups. 122

123

Figure S27: De novo clustering of the validation set and prediction of the discovery set. a. Plot of the adjusted Rand index over all possible k values indicates that k = 10 is a global optimum when linear discriminant analysis (LDA) is used to predict subgroup assignment in the discovery (D1) cohort based on the validation cohort (D2) and vice-versa. Here the adjusted Rand index is used to assess predictive accuracy. b. Genome-wide gene-centric frequencies (F) of somatic copy number aberrations (CBS-derived) are illustrated for each of the integrative clusters discovered in the validation. Note that the cluster numbering does not correspond to the IntClust groups discovered in the discovery set. c. Genome-wide gene-centric frequencies (F) of somatic copy number aberrations (CBS-derived) are illustrated for each of the integrative clusters predicted in the discovery set based on de novo clustering of the validation set. Note that the cluster numbering does not correspond to the IntClust groups discovered in the discovery set. d. Kaplan-Meier plots of disease-specific survival for the validation cohort and discovery cohort (predicted by the validation cohort). e. The in-group proportion (IGP) and associated P-value (based on 10,000 permutations) indicate the quality of each cluster reproduced in the discovery set. 126

Figure S28: The integrative cluster groups have distinct copy number profiles. Genome-wide gene-centric frequencies (F) of somatic copy number aberrations (HMM-derived) are illustrated for each of the integrative clusters as is the strength of association between an aberration and a particular subtype (-log10 P-value) based on a χ2 test of independence. Subgroups have been ordered by the similarity (based on hierarchical clustering) of their copy number profiles. 127

Figure S29: Genome-wide plot of Spearman correlations between CBS-derived CNAs and expression in cis. a. Genome-wide frequencies and cis-acting profiles are illustrated for all tumours, where black points indicate significant correlations (ρ > 0.5 and Bonferroni adjusted P-value < 0.05). b. As in (a), cis-acting profiles are illustrated for each of the integrative subtypes.

129

Figure S30: Genome-wide plot of Spearman-correlations between HMM-derived CNAs and expression in cis. a. Genome-wide frequencies and cis-acting profiles are illustrated for all tumours, where black points indicate significant correlations (ρ > 0.5 and Bonferroni adjusted P-value < 0.05). b. As in (a), cis-acting profiles are illustrated for each of the integrative subtypes.

131

Figure S31: Unsupervised clustering of genome-wide copy number data in 997 tumours from the discovery set indicates heterogeneity within the intrinsic subtypes. Samples were sorted according to their intrinsic subtype classification and clustered based on the copy number profiles (CBS-derived) for 2,995 merged regions using the Euclidean distance metric and the Ward method (red, gain; blue, loss). Genomic instability (GI) based on the proportion of genome altered (black line) and jump measure (red line) is indicated.

132

Figure S32: PAM50 subtype-specific copy number profiles for CBS.

133

Figure S33: PAM50 subtype-specific copy number profiles for HMM.

134

Figure S34: Kaplan-Meier plot of disease-specific survival for the integrative subgroups in the validation cohort.

135

Figure S35: Multivariable Cox regression analysis and Kaplan-Meier plots of overall survival for the integrative subtypes. a. Graphical representation of the Cox proportional hazard ratios for overall survival in the discovery cohort for several key variables including, grade, tumour size, age at diagnosis, number of lymph nodes positive, as well as the integrative cluster subgroups. The hazard ratio is illustrated for selected values of the covariates, as indicated where each subgroup was compared against IntClust 3. Confidence levels correspond to those indicated in the legend. b. As in (b), but for the validation cohort. c. Kaplan-Meier plots of overall survival for the integrative cluster subgroups in the discovery cohort. d. Kaplan-Meier plots of overall survival for the integrative cluster subgroups in the validation cohort. 136

Figure S36: Subtype-specific gene networks are revealed by the CNA-expression landscape. The CNAdevoid subgroup (IntClust 4) exhibits a strong adaptive immune and inflammatory response, which is driven by a trans-acting TCR deletion signature. Genes common to the immune response signature and that are modulated in trans by TCR deletion events are outlined in green. Up-regulated genes are indicated in red, whereas those that are down-regulated are shown in blue.

137

a

T−statistics from contrast absent vs mild or severe LI

Density

TCR-γ (Chr7) assoc. TCR-α (Chr14) assoc.

−5

0

5

10

T Statistic

b

T−statistics from contrast severe vs absent or mild LI

Density

TCR-γ (Chr7) assoc. TCR-α (Chr14) assoc.

−4

−2

0

2

4

T Statistic

Figure S37: Histograms illustrating the enrichment of TRG and TRA trans-associated mRNAs amongst genes that were differentially expressed in the CNA-devoid subgroup according to the extent of lymphocytic infiltration. a. Histograms illustrating the distribution of T-statistics for mRNAs associated in trans with the TCR deletion events in CNA-devoid cases with no LI versus those with mild or severe LI. Genes associated with the TRG deletion on chromosome 7 are indicated in orange and the TRA deletion on chromosome 14 are indicated in turquoise. b. As in (a), but for the comparison of severe LI versus absent or mild LI.

138

Figure S38: Subtype-specific gene networks are revealed by the CNA-expression landscape. The Basal subgroup (IntClust 10) is characterised by cell cycle and DNA repair processes that largely overlap with a Basal-specific chromosome 5 trans-acting deletion. Genes common to the pathway signature and that are modulated in trans by chromosome 5q deletions are outlined in green. Up-regulated genes are indicated in red, whereas those that are down-regulated are shown in blue.

139

Figure S39: Subtype-specific gene networks are revealed in the CNA-expression landscape. The 11q13/14 cis-acting, poor prognosis subgroup (IntClust 2) is enriched for genes implicated in the cell cycle and developmental related processes. Up-regulated genes are indicated in red, whereas those that are down-regulated are shown in blue.

140


doi:10.1038/nature10983

19

Supplementary Table Legends

141



doi:10.1038/nature10983

Table 1: Clinical and demographic characteristics of the discovery and validation tumour cohorts

Table 2: Clinical annotation for the discovery cohort of 997 breast cancer patients

Table 3: Clinical annotation for the validation cohort of 995 breast cancer patients

Table 4: Gene-centric frequencies of somatic aberrations (derived from CBS) across various subtypes

Table 5: Minimal common regions of alteration for CNAs (derived from CBS), stratified by ER status and PAM50 subtype for regions with frequency > 0.05

Table 6: Broken gene frequency summarized across CBS-derived copy number changes, where the pattern of breakage is indicated along with the total number of samples in which a break was observed

Table 7: Gene-centric frequencies of somatic aberrations (derived from HMM) across various subtypes

Table 8: Broken gene frequency summarized across HMM-derived copy number changes, where the total number of samples in which a break was observed is indicated

142



doi:10.1038/nature10983

Table 9: Verification of germline CNVs and somatic homozygous deletions

Table 10: List of primer sequences for verification of homozygous deletions

Table 11: Comparison of CNA calling accuracy between CBS and HMM based on MLPA and corresponding frequencies of alteration

Table 12: Minimal common regions of alteration for germline CNVs (derived from CBS), stratified by ER status and PAM50 subtype for regions with frequency > 0.05

Table 13: Probe-level regions of germline CNVs (derived from CBS) across various cohorts

Table 14: Gene-centric frequencies of germline CNVs (derived from CBS) across various subtypes

Table 15: Distribution and genomic locations of germline CNVs (derived from CBS) in normal breast tissue, tumour samples, and the HapMap population

Table 16: Inferred ethnicity based on projection of tumour genotype data onto HapMap population clusters

143



doi:10.1038/nature10983

Table 17: Genome-wide associations of tumour expression with SNPs, CNVs, and CNAs (derived from CBS)

Table 18: Genome-wide associations of tumour expression with SNPs, CNVs, and CNAs (derived from CBS) in cases of European ancestry for which normal genotype calls were available

Table 19: Accuracy of genotyping calls in non-diploid copy number regions

Table 20: Genome-wide associations of normal expression with SNPs, CNVs, and CNAs (derived from CBS) in cases of European ancestry for which both normal genotype calls and normal gene expression profiles were available

Table 21: Mean and median variance explained by SNP, CNV, and CNA tumour expression associations

Table 22: Genomic regions exhibiting high-level amplification, amplification and putative homozygous deletions (derived from HMM) driving expression to the extreme tails of the population-level distribution

Table 23: High-level amplifications driving expression to the extreme tails of the population-level distribution (derived from HMM) and the frequency of the derivative copy number event for both HMM and CBS analysis

Table 24: Putative homozygous deletions driving expression to the extreme tails of the population-level distribution (derived from HMM) and the frequency of the derivative copy number event for both HMM and CBS analysis

Table 25: Enrichment analysis of trans-acting aberration hotspots. Lists of trans-acting aberration hotspots (denoted by CNA_ID), their cognate mRNAs (denoted by GENE_NAME), and the -log10 P-value of association between a copy number aberration and gene for varying window sizes

Table 26: Enrichment analysis of trans-associated mRNAs for the 10Mb window, labelled according to the corresponding CNA_ID in the aberration hotspot file (Table S25)

144



doi:10.1038/nature10983

Table 27: Pathway enrichment of genes with mRNAs correlated with deletion of the TRG locus Table 28: Pathway enrichment of genes with mRNAs correlated with deletion of the TRA locus Table 29: Pathway enrichment of genes with mRNAs correlated with deletion of chromosome 5q Table 30: Gene-centric ANOVA to determine copy number associated gene expression changes in cis for CBS Table 31: List of gene expression and copy number features input and selected for the k = 10 partitioning based on integrative clustering of CBS-derived copy number changes Table 32: List of outlying features used as input for K-Means clustering on expression data Table 33: Cox proportional hazard ratios for a. clusters based on the de novo clustering of the validation cohort and b. clusters predicted in the discovery cohort from the validation cohort Table 34: Gene-centric Spearman correlation analysis to determine copy number induced gene expression changes in cis amongst the integrative subtypes (CBS) Table 35: Gene-centric Spearman correlation analysis to determine copy number induced gene expression changes in cis amongst the integrative subtypes (HMM) Table 36: Gene-centric analysis to determine copy number associated gene expression changes based on both non-parametric (Kruskal-Wallis rank sum test and Wilcoxon rank sum test) and parametric (ANOVA, Tukey honest significant difference) tests (CBS) Table 37: Gene-centric analysis to determine copy number associated gene expression changes based on both non-parametric (Kruskal-Wallis rank sum test and Wilcoxon rank sum test) and parametric (ANOVA, Tukey honest significant difference) tests (HMM) Table 38: Gene-centric χ2 analysis to determine subtype-specific copy number changes amongst the integrative subtypes (CBS) Table 39: Gene-centric χ2 analysis to determine subtype-specific copy number changes amongst the integrative subtypes (HMM)

145



doi:10.1038/nature10983

Table 40: Cox proportional hazard ratios for a. the discovery set, b. the predicted memberships of the validation set, and c. a comparison of the validation set relative to the discovery set

Table 41: Comparison of C-index values for the IntClust model, PAM50 model, and clinico-pathological model

Table 42: Centroids based on the nearest-shrunken centroids method for classifying samples according to the 10 integrative clusters

Table 43: Matrix of features from integrative clustering for the discovery and validation set used for classification

Table 44: Predicted cluster membership for the validation cohort based on PAMR

Table 45: Scoring of lymphocytic infiltration in the CNA-devoid subgroup (IntClust 4)

Table 46: Differentially expressed genes for contrasts comparing each of the integrative subtypes versus normal breast tissue

Table 47: Summary of pathway enrichment amongst the integrative subgroups

146
